AI Techniques Used in PV Systems

Ensemble Learning

It is a type of machine learning methodology designed to improve the ability of generalization in learning models using several learners. This learning scheme is suitable for small data models. It comprises steps like boosting, bagging, and stacking. In boosting and bagging, homogeneous learners are involved, whereas heterogeneous learners are involved in stacking stage. In Ref. [5], the various ensemble-learning approaches are applied to calculate the efficiency of power conversion of organic dye-sensitized solar cells. This methodology exhibited its capability in the exploration of complex quantifiable structure activity connection for the case when features are distant from targets. Ensemble methodology outperformed methods such as support vector machine (SVM) and single base learner method and was able to achieve excellent generalization and higher accuracy. In Ref. [6], data-based ensemble methods have been used in predicting the generation of solar energy. This methodology has shown ability to handle the intermittent behaviour of generation due to PV systems. A one-day-ahead forecast of solar power generation was improved using optimized an Artificial Neural Network (ANN)-based ensemble method. Bagging and trial-error process was used for the optimization of hidden neurons in ANN model. Further, the bootstrap method was embedded into the ensemble for the estimation of the uncertainty in sources that may affect the predictions of models. A real case study was carried out on a grid-tied PV system (231 kW) for forecasting and showed that the ensemble approach-based model outclasses other benchmark models such as persistence model and optimum ANN model. The benefits of the proposed scheme were accurate short-term prediction of power generation, efficient scheduling, operational contribution of power generation in energy mix, trustworthy operation, and economic benefits. In Ref. [7], an improvised form of ensemble learning aimed at improving the forecast of power generation from the PV system is presented. The proposed ensemble model employed a combination of an adaptive residual compensation scheme and an evolutionary algorithm.

The experimental results showed that the proposed technique performed better than other conventional techniques such as least-square boosting ensemble and weighted average ensemble techniques. In PV systems, accurate forecasting of solar power is essential. There are various data-based methods like ANN, SVM, learning machines, boosting regression trees, etc. proposed for the accurate prediction of solar power generation. However, each of these methods predicts with different accuracies. Therefore, in order to find the accurate prediction, ensemble method can combine the predictions of different models and can find the mean while enhancing the accuracy of predictions.

Deep Learning

It is termed as one of the best technique among machine learning techniques and finds use in the various following applications:

• • Processing of audio or speech
• • Machine translation
• • Computer vision
• • Filtering contents of social network
• • Big data applications
• • Solution of Al-based problems

In renewable energy systems (RES), it is desired to develop an accurate prediction model. AI techniques have proven their role as a tool for optimizing and predicting the power generation in RES. The challenges found in RES are analysis of large data and processing of larger number of variables. In order to tackle these problems, use of deep learning (DL) algorithms is recommended by researchers. Some of the proposed DL techniques in recent years are, namely, convolution neural network, Boltzmann machine, and auto encoder.

Convolutional Neural Network (CNN)

This methodology consists of various layers of neurons that are trained to achieve a high performance. It comprises training in two stages, namely feed-forward and backpropagation. In the beginning, convolution operations were performed on input targets and the parameters of each neuron. The output of network is then compared with loss function, which yields error. This computed error acts as an input for next stage, i.e. backpropagation. At this stage, the parameter’s gradient is computed and each parameter is altered and fed-forward. After some iterations, the training of the network ends.

Boltzmann Machine (BM)

These methods are based on stochastic ANN used to train probability distribution. Restricted BM and deep BM are popular among BM techniques. In restricted BM, a limitation is introduced by constructing a bipartite plot using hidden and visible units. Both of these units are independent, with some constraint. This feature offers the evolution of optimized algorithms for training. In deep BMs, there are several hidden neuron layers with individual layers of even number. During the training of layers in an unsupervised model, self-managed learning (SML) algorithm is employed to optimize the probability distribution. In literature, survey was done on deep BM technique for predicting the output of a solar farm in Germany. Data was collected for 1000days (approx.) with a resolution of 3 hours per day.

Auto Encoder

Auto-encoder DL techniques incorporate the learning of encoders, which can reconstruct the input, and the corresponding output vectors also have dimensions of the input vector [8]. Optimization of encoders is done to keep the reconstruction error at a minimum. From the previous encoders, a code is learned and sent to the next autoencoders, which are trained using backpropagation. This method is quite efficient among DL techniques.

In Ref. [9], various DL techniques implemented in PV systems are reviewed. They also proposed a new taxonomy for evaluation of the performance of PV systems. The main objectives of the review were as follows:

a. Investigation of energy policy used with AI methodologies in PV systems

b. Analysis of hybrid and single DL

c. Comparison between single DL/hybrid DL with other computational intelligence methods

On the basis of this review, the optimization techniques-based DL methods have been recommended for the prediction of parameters in PV systems.

Machine Learning

Learning search algorithm (LSA) is inspired by the learning behaviour of disciples. The algorithm begins with a group of random learners. At this stage, students are classified among groups, namely best learner, worst learners, and average learners, on the basis of their knowledge. In the next stage of positive learning, the learner groups increase their knowledge depth. In the final stage, learners are subjected to a negative learning pattern where the best attributes of the worst learners are learned by the other learners. In this manner, the knowledge level of entire group can be enhanced. Support vector machine (SVM) is also a promising machine learning technique aimed at minimization of reconstruction error and maximization of separate margin for various classes. It is widely employed in supervised learning as a regression tool.

In Ref. [10], an ML technique was adopted for the evaluation of output electrical energy of PV inverters using five-year data under both partial shading and unshaded conditions. Prediction of the generated energy of PV systems while using long-term data analysis through the ML technique was done in this work. It was concluded that the average degradation of microinverters was at 3% per year, and no significant variation was found in the annual energy yield of microinverter-based PV systems. In Ref. [11], a sliding mode control (SMC) and reinforcement learning (RL) for a three- phase grid-integrated PV system are presented. The RL-based maximum power point tracking (MPPT) algorithm was implemented and SMC technique was used for reference current generation. Comparison between SMC.-RL MPPT and fuzzy logic-based SMC was done. It was concluded that the SMC-RL technique performed better in terms of extraction and control of maximum power as compared to the fuzzy-based SMC method.