Energy, or more specifically electricity, is one of the most significant pillars of society. Not only it is of vital importance to people’s daily lives, but due to the growing population and economic growth its consumption will continue to increase substantially in the future ^[1][2][1,2]. The changing landscape of energy systems and the increasing dependence on electricity make it necessary to develop strategies to mitigate the impact of energy disruptions [3].

As global efforts to reduce greenhouse gas emissions and decarbonisation continue, renewable energy sources such as solar and wind power are being integrated into the energy systems faster than any other fuel in history ^[1][4][1,4].

Solar Photovoltaic energy has emerged in the last few decades as the most flourishing source of power generation [5]. Not only is it a clean and renewable energy, but it is also economically accessible with minimal maintenance.

Nevertheless, they have the disadvantage of high dependence on climatic factors, significant variability and high cost of energy storage. Hence, forecasting the generation of Photovoltaic (PV) installations for a given period of time can help to make optimal use of resources, allowing for reduced emissions, lower costs, safe operation and better integration into the grid ^[6][7][6,7].

Furthermore predicting solar energy generation offers intriguing prospects and simultaneous challenges as an accurate forecasts enables efficient grid integration and informed decision-making in energy trading and storage [8].

The prediction of PV power generation has been extensively studied in the literature using different approaches; generally a single location is used ^{[9][10][11][12]}[9,10,11,12].

Previous generation values or the irradiance ^[9][13][9,13] at the time are typically used as the only or most important characteristics for predicting PV generation. The problem with this is the lack of a record of this data and its difficult accessibility. For this reason, a large number of studies concentrate on the prediction of irradiance ^{[14][15][16][17][18][19][20]}[14,15,16,17,18,19,20].

In particular, irradiance is a parameter that, depending on the location, may be very accessible or not recorded at all. In addition, it requires precise instruments for its measurement, calibration and maintenance, which generates uncertainty about the reliability of the data.

A number of scientific works have been dedicated to predicting the efficient use of solar panels and the electricity they generate due to the fact that new energy sources are increasingly important for our contemporary society. This section summarises the research carried out over the past few years and the results achieved.

Kim et al. ^[21][22] propose a model for predicting the solar power, obtained from Photovoltaic (PV) panels and for optimising the tilt angle in the case of Daegu city in South Korea. For this purpose, the researcheuthors apply several Machine Learning algorithms like: Linear Regression (LR), Random Forest (RF), Gradient Boosting (GB), Support Vector Machine (SVM), Least Absolute Shrinkage and Selection Operator (LASSO), and also consider several influential factors (weather conditions, availability of dust and aerosol).

It has been proven that such an approach leads to increased effectiveness at energy production. Wei ^[22][23] investigates how to improve the functionality of PV systems, which are located in Tainan City, Taiwan. Solar radiation of the panels’ surface at various tilt angles is predicted via the utilisation of four Machine Learning (ML) algorithms: LR, RF, Multilayer Perceptron (MLP) and K-nearest Neighbors (kNN). The optimal value of the solar panels’ tilt angel is also found. Machine learning techniques related to the construction of Artificial Neural Networks (ANNs) are used by Kamal et al. ^[23][24] to reconfigure the topology of PV arrays and to achieve their optimal workability.

The findings point out that the presented mechanism with very high accuracy is capable of to outlining the best topologies (among the following: series parallel, parallel, bridge link, honeycomb, and total cross tied) for the PV panels’ deployment. Dependence between the correct PV panels’ installation and their efficiency is investigated by Kim and Byun ^[24][25] as the researcheuthors predict power generation. The XGBoost algorithm is applied for solving a regression task and to give a very accurate prognosis regarding the electricity generation. Khilar et al. ^[25][26] propose a model, based on the deep belief network, for detecting the dust level on solar panels. Such an investigation is important for the places where there is almost no wind and rain, and at same time, the PV system must work efficiently. The considered input variables include dust particles on panels, temperature, and solar irradiance, which are important for identifying the frequency of the manual or automatic cleaning of the panels.

Khan et al. ^[26][27] rely on ensemble ML algorithms RF, XGBoost, and catboost to find the optimal direction for the placement of PV panels. The proposed approach can predict solar power at two levels: at the first level, base models are created via the utilisation of XGBoost and catboost, and the resulted predictions are used at the second level where RF is applied for building a metamodel. The task of this metamodel is to learn what is the best way to use the predictions gathered from the base models. The presented method is evaluated and compared with other ML algorithms and it is proved its better performance is proven. Gautam M et al. ^[27][28] show a framework to maximise the usage of solar power, which is built on the Decision Tree (DT) algorithm. The idea behind it is to find a strategy for switching solar and grid systems, which are connected to a common node, to perform efficient energy management. This, DT predicts the switching configuration and, in this way, the cost for electricity is reduced at an increased consumption.

Predicting the usage of solar power in summer and winter in Mashhad, Iran is discussed by Almadhor et al. ^[28][29]. This investigation is conducted in the context of the realisation of smart cities and through stimulating the citizens to used renewable energy. The prediction is performed via constructed ANN, which solves a linear regression problem. Shaaban et al. ^[29][30] use the Machine Learning approach to evaluate the dust volume on solar Photovoltaic panels as dust contributes to decreasing the generated energy. The proposed model, based on a regression tree, is compared with an ANN model in order to be demonstrated its high performance. When a threshold value of the dust level is reached, a cleaning procedure is triggered. Bulusu et al. ^[30][31] propose a predictive model that points out the hourly energy production from solar microgrids. This novel approach uses ANNs and includes two parts: (1) for features extraction and (2) for predicting the energy production via the created models for the given hours. The researcheuthors recommend not using data older than two years to train the models, as solar panels degrade over time and significant differences in the training and testing data can occure. An investigation related to the performance of PV grid-based systems is presented by Yar et al. ^[31][32] as the analysis relies on several methods like: conducting experiments, accomplishing simulations, and applying Machine Learning. The Logistical Regression technique is utilised to evaluate the difference in the performance between simulation results and real-time systems. Then, the findings obtained regarding monoperk and polyperk crystalline systems as well as their advantages and bottlenecks are discussed.

Mahesh et al. ^[32][33] propose a novel method for evaluating the efficiency of PV panel systems. It combines the Maximum Power Point Tracking (MPPT) technique and SVM Machine Learning to predict the maximal value of the power generated by a PV solar panel. The presented method is compared with the available ones (ANN, fuzzy logic, perturb and observe, and incremental conductance) and its high performance is proven. The problem regarding how power effective the PV panels placed on the building’s façade are examined by Vahdatikhaki et el. ^[33][34]. Surrogate modeling is applied to simulate solar radiation as the last one is predicted via the RF algorithm in three scenarios. Then, an optimisation procedure is applied regarding the obtained RF model hyperparameters via the usage of the genetic algorithm. The researcheuthors conclude that such an approach possesses big potential to simulate and predict with high accuracy the solar radiation of vertically placed PV panels on the buildings’ surface.

Pasion et al. ^[34][21] uses Machine Learning techniques to construct models based on data from twelve sites in the United States of America to predict Photovoltaic energy production without irradiance data. Incorporating irradiance data into solar energy forecasts poses challenges in terms of data accuracy, computational complexity, maintenance costs, lower interpretability and limitations in regions with sparse historical data. The researchtudy uses readily available parameters such as location, time and weather conditions. By comparing six Machine Learning algorithms, including deep learning and ensemble models, the researchstudy identified distributed random forest as the most effective. The researchstudy also found that ambient temperature, humidity and cloud cover were the most important variables for prediction. The researcheauthors highlight the possibility of precise predictions even without irradiance data.