2.2.1 Prediction of Groundwater Level

The research [18] aimed to forecast groundwater levels based on precipitation and temperature, on different timings. It uses the multiple layer-based perception model, genetic programming, and the Radial Basis Function Neural Network (RBFN) which is the Whale Optimization Algorithm (WOA) model. The MLP model is designed and tested with the multiple inputs with the delay of 3, 6, and 9 months to obtain the best results.

Ground level water is part of most significant resources of water supply. But, in recent years, it has dropped due to inappropriate mining and soil depletion, resulting in shortage of water in present and also in future [18, 19]. In recent years the usage of soft computing groundwater forecast models such as ANN has seen a rapid growth in groundwater forecasting [20, 21], for preparing itself structurally for the prediction [22]. Using factors such as bias, weights interconnected, volume of hidden layers, and neurons, ANN model demands that its structure parameters should exhibit precise determination [23]. Among ANN model, the WOA-based model is the most powerful ones because of its various characteristics to pair with other algorithms. Further, optimization algorithms can be used to resolve issues with wide number of variables for the decisions. So, it is widely used in numerous fields like addressing the optimal control issues [24], improving the power infrastructure [25] and engines in vehicles [26].

WOA [27] has been proposed based on the hunt humpback whale activity that enhances the neural network by recognizing the effective input data using the correlation method with the numerous time delays, undergoing the normalization and training stage on that data, which later controls the stop criteria simulations depending on the error threshold. If the stop criteria are appropriate, then it executes the test stage and demonstrates the production results. If the stop criteria are not appropriate, then it undergoes different steps to pick parameters for neural network perception. The parameters also include kernel distance, neuron content width, and quantity of weights as the variables of decision and whale’s primary position to get the production results. The genetic programming prediction which is the approach based on the three-formula structure that includes the mathematical formulations like addition and multiplication, with the terminals of constant numbers and problem variables. This model is used for predicting the variables value. The genetic programming model is one among the most significant hydrological variables prediction model.

Yazd is Iranian province which is struggling from aridity and drought, due to its exotic climatic conditions. The Yazd-Ardakan desert area [27] is one of the Yazd region’s biggest water reservoirs having 1,085 depth wells with the water discharge rate of 260,182 m³/year, 255 semi-depth wells with a discharge rate of 19,764 m³/year, 827 ducts with a discharge rate of 104,124 m³/year. At the present, the rate of exploitation in Yazd-Ardakan desert is 515.5 million m³ with 132 million m³ balanced ground water, that is dropping drastically which is resulting in the change of aquifers width level in different regions. Anticipating aquifer level plays a major role. The data for various weather conditions in different time intervals were collected for the period between 2000 and 2012 in the Yazd-Ardakan desert area. These data were fed as an input to the modeling process at different stages using genetic, whale, and firefly algorithm for predicting the water level, in which the WOA was completely successful.

The WOA [28] model’s results obtained were successful in the prediction of ground water level. The foresight executed will enable planners prepare better for the development of water management in areas of high quality [27]. Basically, the forecasting analytics are valuable instruments for showing the dropping groundwater during times of drought. Integrating the pre-processed data with the soft computing-based model will decide the groundwater duration and value of ground water level.

As drought conditions are harder to define in terms of time and in space, they are associated to water shortage precipitation, river flows, soil humidity, or a composite of all three [29, 30]. Drought is not an immediate occurrence but a slow response of hydrological cycle elements to scarcity of rain which makes it important to minimize its effect along the effective decision-making systems. Agricultural water management is becoming drastically high in case of limited water availability caused by drought, since the agriculture uses maximum quantity of water when compared to all other sectors. If the agriculture sector incorporates the efficient method of using the water, then most amount of water can be saved for other sectors. For this, the knowledge about the relationship of crop yield with the usage of water gains an important role [31–33]. The nations with long-term drought strategies are better equipped to cope with droughts than the nations with small term strategies [34–37]. So, it is necessary to have effective management system incorporated with prediction and smart alert [33, 34].

In the region of Ethiopia, the occurrence of the drought is highly normal, specifically in the regions of the Awash River Basin [35, 36], so far, no advance warning system has been introduced there. So, the long-term model was developed to forecast the stream flow for the irrigation in Middle Awash Agricultural Development Enterprise (MAADE) [37, 38] in the regions of Awash River Basin of Ethiopia. The design has been used to suggest effective management plans for agricultural water. It is implemented during the phases of drought-induced water deficit to reduce the effect of water scarcity. Strategies proposed were entirely focused on surface-based irrigation, where water from the river is supplied directly to the stream canal diversion. This method can be used by the water managers for the effective planning of water management which is based on the water supply forecasting [37].

The Awash River stream flow forecasting model was developed using with a three-layer back-propagation network. Statistics on weekly time series (1987–2001) was used to implement a long-term stream flow prediction model based on the historical data at the MelkaSedi stream gauge station [37]. The model developed was improved using the 10.5 years data and was verified with evidence from the next 3.5 years (1997–1999/2000). So, whenever the water falls below the predefined threshold level in the river that will be reported by the time series predicted for stream flow and thus the effective water management strategy can be adopted in the MAADE irrigation scheme [37].

So, depending on the predictions, different assessments were taken. The assessments were based on various scenarios of MAADE’s agricultural growth strategies such as current planting trends, shifting the year for planting, modified crop varieties, and decreased area of cultivation [39]. Potential options for agricultural production in the system with suitable methods for the control of agricultural water is determined on the basis of residual surviving stream through the river after the water requirement scheme was met [39]. Thus, the implemented water conservation techniques were successful.

In most of the coastal areas ground water is unfit for drinking due to the over-exploitation. So, tracking of groundwater level is extremely significant in this area, which depends on the logical resources that are available and the hydro-geological aspects. The continuous tracking of ground water level is at the most priority in order to manage the groundwater resources in the coastal zone.

The visualization of hydrological variables and the analysis of physical process mainly undergo the important resources like physical and conceptual models [40, 41]. These models contain a massive quantity of advanced, high quality data, measurement systems using systematic methods of optimization, and a thorough knowledge of the relevant physical process [42]. A one-dimensional, conceptual approach of the flowing water would be used to estimate the yearly groundwater with weather-based variables [43]. The modeling of shallow water tables is done through the open-loop with self-exciting threshold autoregressive and using stochastic differential equations [44, 45]. In most of the water table fluctuation modeling, ANN-based models are used. Groundwater level prediction using ANN is framed by continuous monitoring of the individual well and the climatic changes [46, 47].

Kakinada is the Eastern Godavari Headquarters in Andhra Pradesh, located to the east cost of India. It has a tropical weather conditions and often impacted by cyclonic storms and dispersion in Bay of Bengal [48]. The area of this region was taken to study, that is parallel to the cost. The data pertaining to temperature, rainfall, and evaporation were collected and the water level data on monthly basis were collected from the open wells (data from 1995 to 2004). The largest amount of rainfall in this area comes from southwest monsoon between July and October [48]. The groundwater levels in this region are dropping, and this may be due to groundwater demand rises and low rainfall.

The different ANN algorithms were performed for the data obtained to forecast the levels of ground water at one-month lead for a coastal urban aquifer in Kakinada, Andhra Pradesh, India [48]. The algorithms where taken forwarded by the RBFN [49], Bayesian Regularization (BR) [50], and Levenberg-Marquardt (LM). Among these, the LM [51] showed the effective results in individual wells. The ANN platform is only good for making short-term yet numerical predictions modeling fits long-term forecasts. Thus, ANN dependent architectures were a superior option for groundwater prediction where Hydrological Parameter Information is limited. The architecture of RBFN [50] is shown in Figure 2.1.

The models of groundwater emulation have evolved as a significant tool for water resource analysts and developers, to maximize and protect groundwater exploitation. Physically dependent numerical models have been used primarily in groundwater systems modeling and research. If the data is insufficient, then the physically dependent models will not give the effective results. So, empirical models are used in order to produce better outcomes by taking less time and data. The technology of ANN is regarded as universal approximators and is capable of defining a relationship from defined pattern [53]. The ANN has the capacity to learn and generalize from the limited data available that allows them to solve the large-scale problems with much complexity [54] which includes the problems of water management [55].

The study was made on forecasting the level of groundwater in the group of wells on weekly basis. The study was based on the Mahanadi Delta of Odisha region in India. The River Kathajodi and its branch Surua flow in this region [56]. The river island is with a total 35 km² surface area, occupation of the inhabitants as agriculture, where paddy is the main croup of cultivation in monsoon season and other vegetables in the post monsoon season [56]. The research area includes around 100 tube wells by government as a primary source of groundwater. Along with these tube wells, there are also some dug wells in this area which becomes dry during summer seasons, thus creating water scarcity in this region.

The data was collected on weekly basis, related to the level of groundwater between February 2004 and June 2007. The weather data are acquired from nearby weather stations [56]. The feed-forward neural network (FNN) is designed, keeping the obtained data as the base used effectively for modeling and predicting variable water sources [57] simultaneously. GDX [58] is used to predict level of groundwater in a broad number of wells. Simulation results shows relatively good ground water level prediction, while accuracy of prediction is found to decline with growth in lead time. This resulted in a week advance groundwater prediction [56] as well as helping in effective groundwater utilization and planning.

Water systems use estimated modeling of long-range demand of water to build their systems and prepare potential plans for water needs [59]. As water distribution systems overloads, water services must modify the implementation and control of their current water distribution schemes. They also need to advance the identification and forecasting of water requirements to prevent expensive plans overdesign using the prediction models through the effective AI techniques [60]. The daily water demand will be very high per day, so the average rainfall and air temperature is considered for building the AI-based neural network models which can be termed as ANN. An ANN is a fully integrated network with several basic processing operations blocks which are highly used in water-based modeling [61, 62]. The widely employed ANN in engineering disciplines is a back-propagation-based [63] ANN model due to repeated training [63, 64] to minimize error.

The Kentucky American Water Co. (KAWC), the biggest water distribution company provides daily water to Lexington, KY, and nearby areas requests Lexington data for 11-year duration [65]. The data collected were used for regression models that are forecasted based on the weekly water demand [65] (Figure 2.2). The back-propagation model is developed from the data with the two-rule–based expert system models, characterized by three notable patterns that are historical water data request, rainfall, and the peak air temperature per day.

In this research, eight model structures were built to daily water forecasting, the first four models were developed using the conventional modeling with time and regression series and other four for modern AI methods [65]. Water consumption usage, peak humidity on the daily bases from total rainfall data of Lexington (1982–1992) are used to develop and test the models. The research shows that the comparatively recent team of advanced ANN techniques provides a better system for short-term water management prediction than other modern approaches [65].

Ground level water is growing as important resource of the nature and integral part of the reliable water supply in all the regions around the world [66], as it can be supplied quite cost-effectively, quickly, and conveniently to deprived communities compared to surface water [67]. But due to increased population, rise in water contamination, and biodiversity, the ground water has got the negative effect which leads to imbalance in the supply of water and other water habitats [68, 69]. Thus, the groundwater planning gained its importance [70] for existing and future decades using physical and ANN models. The most supportive system for this planning is ANN models [71] because the physical models are costly and require more labors and data intensive [72], whereas ANN does not specifically require the characteristic of the surveillance system which makes them cost-effective and able to learn from the past data that have been collected. The ANN models were developed with the collection of input variables, established for the prediction of ground level water in a single or few well. The triangle algorithms are mainly employed in implementing groundwater level prediction ANN, notably BR, LM, and GDX algorithm. The back-propagation algorithm has worked efficiently in many applications, but even that has some limitations, these algorithms are developed to overcome those limitations. The LM algorithm (LMA) will limit the sum of square error and increases the process than back-propagation [73]. Increased hidden layer neurons leads to over fitting issues, resulting in less performance in prediction of both big and small networks. This can be overcome by BR algorithm [74, 75]. The algorithm’s output is really sensible to the right setting on the pace of learning. To overcome this issue, the GDX algorithm is used which is the combination of adaptive learning rate with momentum training.

The Bayalish Mouza in the basin of Kathajodi river of Orissa state, India, is the region selected for research [76]. It is the river island of 35 km² in area, enclosed by river Kathajodi and Surua with the humified topical weather conditions and an average rainfall of 1,535 mm per year. Ground-based water is the primary agricultural source of this region; there are 69 tube wells that are significant groundwater drainage sites [76]. The level of ground water drastically responds to the level of rainfall, change in the stage of the river that strongly reflects the relation to surface-based water body [77]. The water is sufficient during the monsoon season, but there is scarcity in the summer season. Weakly data related to ground water level in this region are collected at 18 sites between February 2004 and June 2007. The collected data is used to train the three ANN algorithms with feed-forward architecture, namely, BR [78, 79], LM [73], and GDX algorithm. The generic representation of feed-forward architecture [80] is depicted in Figure 2.3.

The weekly statistics related to the rainfall, evaporation, channel’s stage, rate of water pumped, and also data related to the level of ground water in the past week are feed to the ANN as input, to form 40 and 18 input-output nodes, respectively [76]. The results of the three ANN algorithms, viz., GDX, LM, and BR algorithm, were tested through visual awareness and statically salient indicators. All the three model’s outcomes were quite similar, but the BR algorithm performance was highly effective [76]. Thus, all the models that were developed can be recommended to forecast the ground level water on weekly basis over the research area, eve for the high lead time wells [76].

2.2.2 Rainfall Modeling

The correlation between rainfall and runoff is one among the most challenging hydrological processes to be understood because of changing watershed parameters. In the modern decades, the ANN are mostly involved in the modeling of rainfall-runoff, as practically any feature can be approximated to an arbitrary level of precision by ANN [81]. Despite a multitude of studies are carried out on rainfall-runoff modeling using ANNs, there are still some concerns that need to be discussed even while using ANN for rainfall-runoff modeling network optimal structure, selection of transfer function or form of network, and development algorithm choice [82]. The RBFN and MLP models are efficient models for the rainfall-runoff modeling, as the specifications for RBFN are calculated from input space directly [83], they use the transfer functions of sigmoid type [84] and could be prepared within a small amount of time. MLP can generate precise forecasting data, but training a suitable MLP [85] can take a very long time and range of parameters need to be generated by the neural network.

The study was made on the Malaprabha river basin, which is a tributary to the right of the river Krishna, India, lies 15 ° 000 to 16 ° 120 North and 74 ° 140 to 76 ° 050 East. It has a drainage area of 515 km² stretching up to Khanapur [86]. The catchment weather is usually dry; the average rainfall in the region is 770 mm per year. The detail study was made on this region based on the data from 1987 to 1991. The ANN was developed with the MLP and RBFNconfigurations for rainfall-runoff model. The data from the first 2 years were employed for network training and the balanced data are employed for validation process [86] where the RBFN show much higher efficiency than the MLP during the validation process. Thus, RBFN appears to be better in rainfall-runoff analysis than MLP.

It is clear from the findings that the efficiency of forecast model depends on the type of network selected [86]. The minimal number of neurons in the MLP hidden layers is resolved after a prolonged repetitive process which yielded in less errors, while large repressors can be placed in the RBFN nodes using the orthogonal least squares (OLS) algorithm [83, 87]. The impact of the study reveals that the generalization characteristics of RBFN nodes in rainfall-runoff modeling are corporately limited with MLP [86].

2.3.1 Solar Power Forecasting

Solar radiation is a key parameter in the study of solar energy. But there are no proper modes measuring the solar radiation. So, at present, the solar radiation is measured using the different climatic parameters that are obtained from satellite and weather station data. Thus, the ANN [88] finds its significance in the prediction of solar radiation, using the data related to the meteorological and climatological. The types of ANN mainly employed are MLP [89] and back-propagation learning algorithm in the prediction of solar radiation [90]. Feed-forward, back-propagation algorithm [91], and adaptive model are used in hourly solar irradiance forecasting [92].

The ANN [93–98] fed with the input will first make the non-linear mapping to represent them in the form of biases and weighs. The basic ANN model will have the input layer that is inputted with the collected data, and then, output layer that sends the final data to the computer and multiple hidden layers. In multilayer neural network [95, 96], weights and transfer or activation function plays an important role. The net feed to active function is weights and input vectors scalar product. The activation function is fed to input and then to the output accordingly [95]. The formation of the ANN is followed by collecting the input data, selecting the architecture, learning algorithms, and the training-test set [97, 98].

The data related to 67 cities in India is collected from the Centre for Wind Energy and Technology (CWET), India, and Atmospheric Science Data Centre (ASDC) at NASA Langley Research Centre [99] that contains the 19 input variables for the period of 12 months [95]. Eight complete months are reserved for training, whereas the remaining months are for testing. Similar designs are pursued for other 67 cities. The MLP with three-layer structure is designed using the feed-forward back-propagation [95, 96]. The normalization method was used to set data and network. The minimum square error and regression values have been achieved. Thus, solar radiation in remote areas has less solar measurement sources.

The solar irradiance is beneficial in the study of balanced energy, atmosphere, thermal effect on buildings, and organization of power plants that operate on renewable source, agricultural, and also assessment of environmental behavior [100–103]. This leads to the growth of applications that are based on solar energy giving rise to increased requirement of accurate and prediction modeling of solar irradiance. This helps in monitoring and optimization of solar energy system operations. The computation models developed solar data forecasting are depending on the model accuracy and inputted data [104]. The ANN models forecasting solar irradiance in hourly basis are developed depending on the irradiance data represented in time series [101]. The solar irradiance ANNs developed were modeled using the three approaches that are based on the meteorological parameters [103–109], past observed data [110–113], and the combination of both the meteorological and past observed data [114–117]. Majority of these models forecast an hour ahead, with the major requirement of input is requirement of large meteorological parameters, and in some cases, the parameters are unavailable [102]. To overcome this issue, MLP network was introduced which can forecast one day in advance for the solar irradiance prediction.

The MLP was developed and tested using the data that was collected from the Italian city Trieste, as an application of forecasting the energy that was produced by Grid Connected Photovoltaic Plants (GCPV) established on the roof of Trieste municipality [118]. The data about the air temperature and solar irradiance was collected from the region from July 1st 2008 to May 23rd 2009 and from November 23rd 2009 to January 24th 2010 for developing MLP. The data collected was divided into training, validation, and test set. The parameters for architecture and training were selected [118]. The model was trained employing K-fold (K =10) for each training set. The developed model was then evaluated employing validation set for individual K-ford (K = 10). The process was repeated for different architectures and selected the suitable efficient architecture, tested it finally with the test set. The selection of the architecture depends on the activation function and neuron count in the hidden layer. The cross-validation is the technique used for predicting the efficiency of the model over the unseen data as feature yet [106].

A 24-hour ahead predicting MLP network was developed for solar irradiance forecasting. After many modeling, it was found that the best model for solar irradiation forecasting in Trieste was with the three neurons (TðtÞ; GðtÞ, t); one input layer, 11 and 17 neurons; two hidden layers and 24 neurons; and output layer [118]. That result ranges between 98% and 95% correlation coefficients for sunny days and cloudy days, respectively. A review of the power generated by a GCPV plant of 20 kWp and the one predicted using the established MLP-predictor demonstrated the effective forecasting for 4 sunny days (56 hours) [118]. Thus, the method implemented gave good insights in the planning of sustainable renewable systems operation.

The photovoltaic (PV) power systems with large-scale grid recently established around the world due to the evolution of PV technology [119]. The power that is generated by the PV systems heavily reliant on the change in the solar irradiance with the other factors of environment, the unpredictable change in the PV system results may drastically rise the service expenses of the electricity system [117]. So, the major responsibility of the grid operator is to identify the change in the power generated by the PV system and to plan the power of spinning reserve and for grid operations controlling. So, the operator for the transmission system adding online forecasting of power for the PV systems plays a major development of support system for energy trading, independent power production and energy services to deliver energy from different functions like trading of energy, assessment of security, and financial phrasing [119].

There are many researches going on to build the online forecasting model for PV power production. The most recent one is to have the two-stage approach. At first stage, solar irradiance is forecasted in various time scales based on ANNs, fuzzy logic (FL), hybrid systems [120, 121], and Auto Regressive (AR); the past data and meteorological data are employed in building the regression models. The models developed in the recent years are mostly based on the recurrent neural network [122, 123] that uses past historical data [118], hybrid models that is developed by integration of ANNs and wavelet analysis [119], multistage ANN [124], integration of ANN and library of Markov transition matrices [125], two-dimensional model approach [126], combination of special designed training algorithm and a diagonal recurrent wavelet neural network (DRWNN) [127]. At the second stage, the forecasted data that represents the temperature and irradiance are sourced to simulation based commercial PV software’s like HOMER [128, 129], TRNSYS [130, 131], and PVFORM [132, 133]. The outcome of these is the simulated forecast data related to the AC power that is outputted from the PV system per hour. There are also some other models to forecast, that are produced using the weather data [134]. All these modules drastically increased their limitation as the solar PV systems increase from small scale to larger scale.

The recent research for forecasting the hourly data types for the present day and the next day using Fuzzy ART (Adaptive Resonance Theory), that showcases the energy from PV was built to interact the Distributed Intelligent Energy Management System (DIEMS) [135]. This model can forecast the power output only up to level seven, that makes it less accurate [119]. To overcome this limitation, the 24-hour ahead generated power is forecasted using a RBFN [119]. This RBFN will forecast the PV systems output power directly from the meteorological services and its past data. The numerical weather prediction (NWP) input variables are classified using the self-organized map (SOM) that increases the accuracy of forecasting [136]. The values that are predicted and the PV devices operating data are compared to verify the efficiency of the approach [119].

The radial basis function network model was applied on the PV system located in the Renewable Energy Research Centre (RERC), Huazhong University of Science and Technology (HUST) [119]. It can produce the highest power of 18 kW. The data related to the amount of power delivered to the grid in hourly basis (from December 19th 2006 to December 25th 2006) by PV system that is represented by supervisory control and data acquisition (SCADA) systems was collected along with the other climatically data like speed and direction of the wind, humidity, temperature, and pressure of the air, cloud, solar irradiance, and sunshine [119]. The radial basis function network model was developed from the data collected, resulting in the model output as the 24-hour ahead output power of the system. After many simulations, the excellent results were obtained from modeling having input layer and output layer with 6 and 24 neurons, respectively. These are having 11, 15, and 15 hidden nodes in the hidden neuron and 0.4, 0.3, and 0.1 spread values in three models (sunny, cloudy, and rainy), respectively [119]. The efficient results obtained from the proposed model shows its significance in using it for PV power systems output power forecasting, thus enabling effective planning of operation involved in PV system [119].

Solar had gained its attention due to its non-carbon generation property, and named as source of green energy. In recent years, the advancement of PV generation has led in the growth of capability of installation from 34% to 82% over history, for the Economic Co-operation and Development Organization [137]. But due to the change in the energy sources, it is difficult to determine the power that is outputted from the PV. So, the uncontrolled, non-deterministic output behavior of the PV plant leads in degradation of system by lowering the efficiency [138]. So, the best solution for this condition is to have a good prediction method for the output power of the large-scale PV system connected to grid that leads to the efficient operation of the system.

The prediction in PV system using neural network [139, 140] includes the neural network with the combination of fuzzy rules for solar irradiation forecasting [141]; integration of neural network and wavelet transform for hour ahead PV output power prediction [142], recurrent WNN for solar irradiation prediction [143], firefly algorithm, and fuzzy adaptive resonance theory mapping network with wavelet transform is used for prediction of power in solar [144]. Although there are several methods, each method has their own limitations, which affects the accuracy of the prediction of solar generation. This increases the demand for efficient prediction model. So, the new module that is the advanced version of the Shark Smell Optimization (SSO) algorithm [145] was introduced, called the metaheuristic optimization algorithm that will optimize the PV forecasting’s free parameters, and therefore improving the search ability for both local and global.

The triple stage forecasting model based on neural network with the hydride framework for training with the combination of LM metaheuristic algorithm was developed for PV output prediction which overcomes the limitation of under and over fitting [146]. The prediction model was fed with the temperature, solar radiation and past 24-hour data related to the PV generation as input [142]. The feature selection model [147] with two cascaded filters has been employed for filtering the irrelevant inputs. The three neural networks with multilayer perceptron (MLP) architecture are trained to derive the mapping function for selected input and output vectors, that makes all three neural networks to use each of their weights by the another one. After learning, the neural network 1 will send its weight to neural network 2; the neural network 2 will learn better with more accuracy (for prediction) and send its weights to neural network 3; similarly, the neural network 3 will become more accurate in its efficiency of prediction. So, by increasing the number of neural networks in cascaded structure as shown in Figure 2.4, the efficiency of the model can be increased, that makes models accuracy negligible with respect to last neural network in the cascaded system [146].

The three neural networks are first trained using LMA, even though LMA is quick learning and efficient, it has the problem of local minima trapping leads to objective function’s (OF) gradient as zero. So, after training with LMA, the resulting weights are fed to metaheuristic optimization algorithm. Based on the criterion of early stopping, each neural network’s LMA state is terminated, in order to eliminate the issues of overfitting [148, 149]. The resulting neural network weights will be fed to improved SSO (ISSO) algorithm; the ISSO’s has error of the neural network training, and thus, the local minimum problems of LMA are solved by ISSO [146].

The model is developed with the combination of two networks; one is the neural network. The other one is ISSO which was further applied on 15KW PV system that is situated in Ashland, Oregon of USA [150]. The efficacy of the metaheuristic algorithm proposed is assessed by correlation with six other algorithms, where the proposed model gave the most efficient results. This model can be used for the photo power forecasting in PV systems [146].

2.4.1 Wind Monitoring

A BPN was developed for checking the performance of a turbine rear bearing by means of SCADA [119] data. The collected information and analogy can be applied for similar turbines. The ANN was implemented for one component and would have been tested in abnormal conditions as well. An unambiguous turbine without gear box was monitored [151]. The technology has come across a gradual transition from fixed speed to variable speed. Pitch controlled has also become feasible to be analyzed in the place of stall controlled. Drive trains with gear boxes are being replaced by with/without gear box. Such transition was made possible with the help of power electronics. Detention of power has proven to be improved with grid and the mechanical stress was also reduced [152].

The convergent steepest descent algorithm has dual parts that include step size and direction of descent. LMA [52] that has Newton’s method and steepest descent algorithm was handed down, which has the advantage of optimizing in a smaller number of steps with better accuracy. Condition Monitoring System (CMS) using ANN can be well explained by Figure 2.5. The imitation of the normal functioning parameters from the monitored data is realized by the ANN and simultaneous application of the detected anomaly. The ANN of single input with 20 neurons and one hidden layer was employed as a regular model. The Nonlinear Autoregressive Exogenous (NARX) ANN that has exogenous inputs was used in which Mahalanobis distance was used as a prediction technique in the anomaly detection. Thus, a turbine can be continuously monitored by which disastrous failure can be evaded 26 hours in advance [153].

Genetic algorithm associated with multilayer autoregressive neural network, FL, and a SIMAP conquered a maintenance calendar based on the technical and financial standards [154]. Only the important aspects of interest would be presented to the operator in a succinct fashion out of large volume of SCADA data to make the fault detection easier [155].

Feed-forward ANN multiple hidden layers were tested for condition monitoring and NARX ANN accomplished the best performance monitoring [156].

2.4.2 Wind Forecasting

The dependency on non-renewable sources has to be reduced. The growing population and existing technologies consume lot of power from nonrenewable sources which will soon result in extinction of non-renewable sources. Our need for electricity can be fulfilled only by renewable energy sources. Power systems can assimilate wind power in large scale by forecasting the speed of wind. Perfectness in the wind speed prediction is significant. The united technique of emplacement of electric power and wind patterns can improve the prediction of speed of wind [157]. MLP training algorithm can be used in the prediction of wind speed. The researchers’ responsiveness toward MLP has increased due to its accomplishments of deep learning. The spatial weights can be added to the input and output weights of MLP, which is well suitable for environmental variables [158].

The 24-hour wind direction, air pressure, and wind speed statistics are collected in hourly basis from March to December. The data was fed to the algorithm for training and the output was resulting the successful prediction of the feature 24-hour wind speed in hourly basis and the wind speed for the next consecutive hours [159].

To forecast the wind speed, feed-forward and feedback ANNs are widely used [160]. Power generation using wind energy is improved in the United States [159]. It is obligatory to have a precise wind speed prediction models that enhances the assortment of locations [161, 162]. The ANN model using the mean squared error criterion for enhancement was tested in Savannah, Georgia. To estimate the wind speed, the ANN model was inputted with longitude, elevation, day, latitude, and min/max of temperature. The results are 95.2% accurate for the feed-forward and 93.2% accurate for feedback neural networks [163].

One of the important tools that can be used in the estimation of linear system parameters is Kalman filter. It necessitates the noise to be Gaussian white noise [164]. In order to spread over this recursive tool for non-linear applications, it needs proper modifications. The Ensemble Kalman Filter (EnKF) is a method to predict wind speed. EnKF linearizes the mean and covariance in the non-linear predictions. Better results are proven to be achieved when EnKF is combined with ANN. Superlative estimation of wind speed by modifying the outputs of ANN using EnKF was one of the preeminent correction structures with better accuracy [165].

The physical, statistical and combined approaches of wind speed forecasting are analyzed and combined method was accepted as a better methodology for the estimation of wind energy. A type-2 fuzzifier was developed to transfigure the nonlinear approach into linear and applied to the MLP which can deal only with linear parameters. The vagueness in the inputs is deliberated by type-2 fuzzifier. The Particle Swarm Optimization (PSO) resolved the ambiguities of the measured parameters from SCADA. Thus, the computational cost has been condensed in the MLP which has only single layer [166]. In WNN, Mexican hat and Morlet wavelets are hidden layer activation functions. WNN is more advantageous over normal Feed-forward networks in such a way that it provides greater simplification due to its adaptive wavelet shapes in conformity to the training statistics [167].

The momentary wind speed was predicted using ANN and the results are altered for the extrapolation of long-term applications with the additive advantages and properties of Markov chains [168]. More accurate prediction of wind power has been achieved by an ANN with LM optimization approach by considering the physical parameters like wind power, pressure, direction, and speed of wind at Jodhpur, Rajasthan in India [169].

A four-layer Recurrent MLP (RMLP) was employed in the prediction of wind power. The network was accomplished with Extended Kalman Filter for training back-propagation through time algorithm. The extremely dynamic altering wind power can be predicted using the combination of RMLP and Extended Kalman Filter [170].

2.5 AI in Geothermal Energy

The main sources of energy are renewable resources that include biomass, solar, wind energy and hydraulic energy [171, 172]. The basic issue in the renewable resources is that the energy from them is not baseload power generation. To make them baseload providers, the fixed amount of energy should be generated from the power stations, with the proper prediction mechanism [173]. Biomass and geothermal are invulnerable to peripheral weather conditions, so they can be used as basic sources of energy [174, 175].

Geothermal as heat energy is transmitted to ground level through hot water and steam. It is created based on the different heat generated inside the crust, atmospheric temperature and also due to the presence of various salts, gases, and minerals [176]. Therefore, the geothermal resources are more related to the temperature. In general, the resources with heavy temperature are employed in production of electricity. The resources with less and medium temperature are employed as direct fields [177]. In the recent research, resources with less temperature are also employed in applications with heat pumps. From 2010 to 2015, globally geothermal power and electricity generation raised 17% and 10%, respectively [174]. The countries like Philippines, USA, Mexico, New Zealand, and Indonesia occupy the top five places in the world in geothermal power and electricity production [177]. The Turkey’s Energy Atlas states that it produces 921.5 MW of total installed power from 32 of geothermal power plants situated there, that contributed 1.2% to the total produced in 2016 [178]. The geothermal power production aims to meet the growing demands of USA that makes the major requirement of feasibility and optimum design configuration for geothermal power plant.

Increase in demand leads in increase of energy cost. So, the analysis of energy to identify the energy losses is necessary. Thus, the exergy diagnosis possesses the important role [179, 180]. The conventional exergy analysis cannot determine the proper dependency and independency between the systems components [177], that effects systems component’s improvement. The recent research has focused on the development and use of advanced exergy analysis models for geothermal power plants [181–183].

The advanced exergy analysis models include ROSENB [184] optimization algorithm, organic ranking cycle (ORC) optimization [185], regenerative ORC optimization, Kalina cycle optimization, and thermal efficiency optimization using ANN and artificial bee colony (ABC) [186–188], supercritical ORC optimization with ANN [189], non-dominated sorting genetic algorithm-II (NSGA-II) for Rankina cycle [190], and PSO [191, 192].

The ABC model [174, 193–196] was developed to optimize the geothermal power plant’s exergy efficiency. The model is assessed by comparing the results of developed model with conventional and advanced analysis methods [174]. By this, geothermal power plant with efficient exergy is designed. The developed model was implemented in geothermal field of Büyük Menderes basin continental rift belt in Turkey [174]. The model was used to optimize and analyze the performance of the binary geothermal power plant with respect to thermodynamical properties. The study suggests that there is a close match in exergy efficiency between advanced and ABC model. It can be concluded that advanced model analysis provides an arbitrary value, whereas ABC model provides a constraint limits that includes minimum and maximum ranges, which is no longer arbitrary [174].

In the recent world, the scarcity of renewable resources is increasing day by day that includes the geothermal energy. The geothermal energy is low temperature resource that generates more than 10% of power, through the geothermal power plants [197]. In the cold regions like Iceland, the 85% heat required is sourced by geothermal energy [198]. The less temperature thermodynamic cycle that is used as the working fluids in organic fluids is called ORC [199–202]. The advanced part of the ORC is Kalina power generation cycle [203–206], that uses ammonia with water as the working fluid, generate power from the unwanted heat [179, 207]. The rate of change in temperature in heat source with respect to working fluid that leads to exergy losses are relatively less in Kalina cycle [208, 209]. The properties of water, ammonia, and their mixture with respect to thermodynamics play a major role in the Kalina cycle [210–212]. The power produced by the kalian cycle through waste heat recovery is greater than the ORC [201]. There are many implementations of Kalina systems like solar Kalina cycle [186], Kalina power system [213], Kalina cycle with waste heat (low-temperature), and LNG cold energy as heat source and heat sink, respectively [214], and Kalina geothermal cycle used in a power plant [215–219]. The continuous monitoring of exergy and energy also plays a major role in the thermodynamic cycles, as the exergy study helps to assess the dropped pressure and heat that is transferred in thermodynamic system [218].

In the recent years, the optimization algorithms are mostly using in the applications of energy systems [219–222]. Using evolutionary algorithms, Kalina and ORC cycles have been commonly optimized to determine the best configuration for thermodynamic cycles [222]. The ABC algorithm [223, 224] is one among the advanced algorithm that helps to achieve the desired thermal and exercise efficiency of the Kalina geothermal cycle. It has advantages in multi-dimensional, multi-objective function, and multimodel issues compared with other popular optimizations techniques [219]. The ABC algorithm [220] as shown in Figure 2.6 has three mains components, namely, food source, employed bees, and the unemployed bees. The unemployed bees are of two types. They are scouts and onlookers [223–226].

The study was carried out on thermodynamic power plant–based geothermal Kalina cycle located in the Husavic power plant, Iceland. The analysis related to thermodynamic is done through the MATLAB. The thermodynamic characteristics of ammonia-water mixtures were measured using the EES software [227]. The ABC algorithm is developed for identifying the optimal exergy and thermal values. Ultimately, the data related to impact of temperature in separator inlet, mass function of ammonia, and working fluid’s flow rate are analyzed [219]. The ABC algorithm has resulted in the behavior of achieving the thermal optimum efficiency. The algorithm may not obtain the appropriate random solution at the initial levels, to find the perfect thermal efficiency [219], but it can in 37 foraging cycles and the optimum thermal efficiency converges gradually. The final optimum thermal efficiency resulted in 20.36% [219].

The geothermal resources utilized for generating the electricity are extracted by boreholes which are deeply drilled in the thermal regions within the geothermal fields [228, 229]. Drilling of this borehole is a complicated task and it affects the initial rock formation around the borehole [230]. After borehole creation, thermal regeneration is assessed by examining shut-in measurements in time and build-up bottom-hole temperature (BHT) [231]. During the process of borehole drilling at various shut in times, the build-up BHT is calculated [232]. The build-up BHT utilizes the sophisticated logging equipment that makes it more costly [233, 234]. In the assessment of geothermal system, critical part is to analyze static formation temperature (SFTs) using the build-up BHT [235, 236].

The analysis of SFS data at early drilling time gives the opportunity to consider the temperature during the creation of virgin, months before the precision calculation [237]. This development helps for the planning, discovery, production and exploitation. The data related to the SFTs helps in many ways like identifying the geothermal gradient for exploration mapping, heat flow in geothermal region, lagging of borehole, difficulty, and design of the borehole and evaluate the properties of forming in situ thermo-physical [237].

The prediction of SFT in boreholes is executed by using the simplified method of analysis, that is based on the complicated heat transform method focused at the bottom-hole circumstances in which BHTs are in fact calculated. The calculations of the BHT appear to show the thermal anomalies [237]. There are many analysis methods like Brennand method [238], Hornerplot method [239], Kutasov-Eppelbaum method [240], Leblanc method [241], and Manetti method [242]. Most of these models require minimum threw or even more observations of BHT in various shut-in times at the same depth of borehole. The SFT identification in this method is done through the mathematical modeling using the input of BHT and shut in time, data related to non-linear and linear part between BHT and each methods time function [243, 244]. Although the effective progress is achieved in recent years in this area, there is a wide difference between the results from the various methods [245]. Therefore, the advanced development of modern, reliable technique to predict SFT is still a difficult task. So, to overcome the complexity [246] in the prediction process, the ANN is developing as a new stable model to measure SFTs in geothermal wells.

ANN as the effective computational technique, applied in various sciences as a modeling tool, addressed many real-world complex issues mainly as a forecasting tool [247]. In the recent research, the application of ANN is found to be increasing drastically in earth sciences [248–252], geothermal, and petroleum engineering [252–256]. The new three-layer ANN was developed to predict the SFT in geothermal wells. It was trained with the database related to the estimated SFT (statistically normal) in geothermal borehole that includes data related to the build-up thermal recovery analyzed during the borehole drilling [237]. The build-up thermal recovery data contains shunt in time and BHT values and transient gradient of temperature. The estimated calculation of SFT value is done through seven analytical methods [237]. The normalization of SFT is done using statistical tests [245], mean SFT, and regular estimated deviations are measured using enhanced outlier identification/statistical process of rejection [257]. The statistical method integrated with ANN approach helps to detect the errors in the result and it also rejects prior to choosing a suitable SFT estimates [237]. The evaluation of the ANN model developed is done using set of four BHU data, for which the SFT was already known. The accuracy assessment of the ANN model was carried out using these validation tests [237].

The new ANN model developed using the multilayer layer and feed-forward architecture has more advantages than analytical method, as it requires BHT and shut-in time as the input mainly [237]. The prediction model is validated by comparing SFT measured and stimulated, that shows the errors less than ±5%. The ANN model proposed can be a functional method for estimating SFT in boreholes. The forecasting accuracy of the ANN model can be increased by using the patterns with more BHT data. The concept of introducing ANN for SFT prediction enables new faster and practical tool for geothermal industry [237].

2.6 Conclusion

This chapter is an overview of the proper utilization of renewable resources like water, solar, wind, and geothermal energy. It also discusses the various forecasting methods using AI approaches in the monitoring, analyses, modeling, and optimization of these renewable sources. The critical findings of this review are summarized as follows.

• WOA has found to be more promising in ground water level forecasting due to its capability of handling many parameters at a time and can be easily paired with other algorithms.
• A back-propagation model with three layers acts as an effective supportive system for drought management.
• LMA gives effective results in ground water level prediction where limited information of hydrological parameters is available.
• GDX method is the most supportive system for long term (1 week) prediction of ground water that helps in effective planning and utilization of groundwater.
• RBFN exhibits better accuracy for rainfall-runoff analysis.
• BR algorithm works well when there is increase of neurons in the hidden layer and performs well for small and big networks.
• Radial basis function networkalong with SOM is found to be the one of the most accurate algorithms in forecasting the output power of PV systems.
• MLP with three-layer structure designed using Feed-Forward Back-Propagation provides better forecasting of solar radiation in remote areas that has less solar measurement sources.
• NARX ANN model is one of the best suitable models for continuous monitoring of wind turbine in which the fault is predicted 26 hours in advance.
• A combined approach of type 2 FNN accompanied by PSO in the SCADA input data of a MLP can be utilized for better estimation of wind power.
• ABC algorithm provides better thermal efficiency in geothermal plants.

This chapter presents a comprehensive work on the restoration of renewable sources using various AI techniques. As a future scope, the research should be focused on more AI algorithms which can provide improved performance methods with enhanced predictions.