Solar plant systems have complex nonlinear dynamics with uncertainties since the system’s parameters and insolation fluctuate [1]. Thereby, it is complicated to approximate these complex dynamics with classical methods, while ML methods provide the required performance [2]. In modern sensor systems, ML methods are crucial units to increase the quality of big dataset processing for solar plant design, forecasting, maintenance, and control ^[1][2][1,2]. Within the EU COVID-19 strategic reply, the smart energy standards define a cloud platform specification for a distributed solar big data ecosystem that will provide the creation of effective ML technologies for smart solar energy [3]. The long-term contribution of solar energy is dependent on overcoming the remaining issues of grid integration, high costs, and low efficiency, mainly through the research and development of a smart solar plant system based on ML methods on account of traditional methods’ ineffectiveness. Within breakthrough studies, ML technologies collected, analyzed, and converted a huge number of sensory datasets into ML knowledge. These big data sets are collected by supervisory control and data acquisition (SCADA) systems [4]. The SCADA system is able to integrate the sensor system and ML technologies into an ML sensor system based on software that implements ML sensor models and integrates with SCADA through API. Further, the application of ML technologies for the digital transformation of solar plant systems has a massive potential to increase their stability, reliability, dynamic response, cost-effectiveness, and other essential advancements, easing their integration into electric grids.

The contribution of this researchticle is threefold. First, researcherswe reviewed more than 100 research papers devoted to state-of-the-art ML technologies of solar plant systems, most of the articles were published in the last five years. Second, researcherswe reviewed resources where researchers can find open datasets, source code, and ML framework and simulation environments to create ML technologies for a solar plant system. Third, in contrast with other review articles, this researchour review proposes a simple but effective pipeline scheme for an ML sensor solar plant system’s implementation and outlines its future digital transformation into a smart solar plant based on integrated, cutting-edge technologies; estimates the impact of the ML technologies based on the proposed scheme on a solar plant value chain; sets benchmarks for performance comparison of the reviewed ML models for a solar plant’s system based on the comparative studies’ results summaries; and briefly summarizes researchers'our self-adaptive models for sizing, forecasting, maintenance, and control of a solar plant based on a modified fuzzy neural net (MFNN) that is automatically created with regard to tasks’ complexities and overfitting problems ^[5][6][7][8][5,6,7,8].

2. Machine Learning Technologies for a Solar Plant’s System

Real-life solar plant systems have complex, nonlinear dynamics due to variations in system parameters and insolation. Thus, ML methods have been proposed to approximate this complex dynamic. The recent studies ^{[1][2][5][6][7][8][9][10][11][12][13]}[1,2,5,6,7,8,9,10,11,12,13] prove that ML technologies for a solar plant’s design, forecasting, maintenance, and control increase the effectiveness and reliability of the solar plant as compared to conventional methods. In smart sensor systems of solar plants, ML methods are crucial units to increase the quality of datasets processing the solar plant’s design, forecasting, maintenance, and control. SCADA is a control system architecture that uses sensors, programmable logic, and discrete PID controllers to control the processes of a solar plant system. The solar plant’s system includes advanced sensors. Big data from SCADA are collected 24/7. Combined with weather big data, this enables the creation of ML technologies to solve complex tasks of a solar plant’s design, forecasting, maintenance, and control.

2.1. ML Sensor System of a Solar Plant

Smart models based on ML technologies have the advantage of parallel computation through modern graphical processing units, which significantly decreases the time cost in SCADA datasets processing for solar plant design, forecasting, maintenance, and control [12]. The reliability, accuracy, and other demanded quality parameters must be composed as the performance of an ML model. This model must be created effectively with high-quality datasets to have optimal performance [14]. Figure 1 shows the basic life cycle of an ML sensor model. Smart model creation has two phases: data preparation (DP phase) and model creation (MC phase). They should be elaborated by the Cross-Industry Standard Process for Data Mining cycle (CRISP-DM) [15] and Open Neural Network Exchange (ONNX) format [16]. The CRISP-DM cycle [15] provides a pipeline for the implementation of smart models in real-time scenarios. The sensor data of a solar plant are compiled into raw SCADA datasets. Then, these datasets are preprocessed (Figure 1) in a simple way (standardization or encoding). Data preparation methods include dimensionality reduction (principal component analysis (PCA)), sampling (subsampling, oversampling), transformation, encoding, feature extraction, and selection [14]. Feature extraction is a crucial step in a smart sensor system’s creation because it provides knowledge for ML model creation [14]. The DM methods generate features. The most relevant data are further separated into train, validating, and test datasets (Figure 1). An ML model to solve either classification or regression tasks is trained based on a train dataset. When a smart model provides the demanded performance, its weights are frozen. The ML frameworks, which rwesearchers review in Section 2.3, provide an automatic MC phase, including validating (Figure 1). The trained ML model is deployed. If a monitored ML model does not provide optimal performance, then it is retrained based on updated datasets.

2.2. ML Methods for Smart Sensor Creation

An ML sensor model can be developed based on neural network (NN) or non-NN algorithms [14]. The last ones include PCA, Random Forest (RF), support vector machine (SVM), and Decision Tree (DT). In contrast with non-NN methods, NN architectures can include various neurons which are specified by ONNX [17], highly effective learning, and extracting features. A deep neural learning/network (DL/DNN), such as a recurrent neural network (RNN), convolutional neural network (CNN), and transformers, is part of the ML methods with feature learning that use multiple layers, complex connectivity architectures, and different transfer operators to automatically mine meta features from the input. NNs, such as artificial neural networks (ANNs), radial basis function neural networks (RBF-NNs), generative adversarial networks (GANs), RNNs, and CNNs have recently made major progress in practical applications of solar energy [1]. Figure 2 shows two NN methods’ classes and the ML method groups according to the task they solved for a solar plant system [2]. The ensemble’s types are bagging, boosting, and stacking/blending ^[18][19][18,19]. Table 1 presents the comparison of ensemble techniques [18]. There are constant and dynamic weighting ensemble approaches. In recent studies, the most used ensemble methods are RF, Extreme Gradient Boosting (XGBoost), Extreme Learning Machine (ELM), etc. Model training methods that optimize performance include quasi-Newton, stochastic gradient descent (SGD), evolutionary computation, genetic programming, etc. [15]. The creation of the ML model is the most complex and important task which includes the creation of an optimal ML model’s architecture and requires a multidimensional global optimization (GO). The bias and variance estimate the effectiveness of a model. The improvement of a model’s bias always makes gains at the expense of variance and vice versa. The performance of ML models highly correlates with the representativeness of a dataset. A lot of techniques provide a model’s evaluation, including cross-validation, kfold, holdout with a different performance including accuracy (ACC), mean squared error (MSE), precision, receiver operating characteristics (ROC), recall, Matthew’s correlation coefficient (MCC), F1, area under the curve (AUC), mean absolute error (MAE), and root-MSE (RMSE). The relative errors, such as normalized RMSE (nRMSE), normalized MAE (nMAE), etc., facilitate the comparison between models that are tuned based on datasets with different scales. With the goal to develop intelligent models for sizing, forecasting, and control of a solar plant system and to make an RNN more adaptive with regard to a task’s complexity and overfitting problem, rwesearchers developed an MFNN ^[5][6][7][8][5,6,7,8]. The MFNN includes RNNs with fuzzy units and/or a convolutional block to process images. An RNN approximates a membership function in contrast to an Adaptive Network-Based Fuzzy Inference System (ANFIS). RWesearchers combined the modified multidimensional quantum-behaved particle swarm optimization (PSO) with the Levenberg–Marquardt algorithm (MD QPSO) and developed a hierarchical encoder of the particle’s dimension component ^[5][6][7][8][5,6,7,8] to automatically create an optimal architecture of an MFNN and improve the convergence. RWesearchers implemented an MFNN and its life cycle, which includes automatic creation and self-adaptation as an intelligent framework based on the researcheauthors’ software [20]. This intelligent framework provides the automatic creation of the optimum architecture of an MFNN with regard to a task’s complexity. All the above-mentioned ML methods and algorithms were implemented as software by an ML Framework, which represents a tool to create a smart sensor system.

2.3. ML Frameworks

ML frameworks implement many ML methods [15]. Table 2 shows the comparison of popular ML frameworks. Big data ecosystems, namely Apache Flink, Apache Spark, and Cloudera Oryx 2, include built-in ML libraries for large-scale DM. These ML libraries evolve presently, but the potency of the entire ecosystem is significant. Google, Facebook, and Microsoft developed most of the DL frameworks that support ONNX, namely PyTorch, TensorFlow, Caffe2, Microsoft CNTK, and MXNet. Chainer, Theano, Deeplearning4, and H2O are also appropriate DL libraries and frameworks for smart sensor system creation. The high-level DL wrapper libraries such as Keras, TensorLayer, and Gluon are developed on top of the DL frameworks. They provide a simpler but more computationally expensive way for smart sensor system creation. The ML frameworks provide an automatic MC phase of an ML model, including validating (Figure 1). An ML sensor system can be implemented as software based on an ML framework that supports ONNX. Such implementation will provide flexibility and all an ML framework’s advantages for a developed ML sensor system.

2.4. Open Resources for ML Research in a Solar Plant System

The open solar energy data sources, including big data, provide the development of cutting-edge ML technologies in solar energy. The GitHub repositories ^[21][22][21,22] are implementations of maximum power point tracking (MPPT) systems [21] and management of cities’ demand/load [22] based on an open-source Gym toolkit [23]. An open-source tool pymgrid [24] provides the creation and simulation of various microgrids. Octave [25] and Scilab [26] are open sources that are compatible with MATLAB. Table 3 presents a brief description of the open datasets to implement and validate ML solar plant systems.