Abstract

The dynamic evolution and variation of electrical loads is now, a priority for their optimal management and, above all, forecasting. Now, these dynamic load variations require computer tools that are able to implement optimal load forecasting models. Scientific research into automated models for forecasting electrical loads is therefore a challenge for scientific researchers, and several studies have been carried out in this area. These include machine learning approaches such as Long Short-Term Memory (LSTM), Support Vector Machine (SVM), Multilayer Perceptron (MLP) and others. These studies are often quite complex due to the number of elevated hyperparameters they contain, with considerable deviations in accuracy between the real and predicted data. Thus, in order to exploit methods with fewer hyperparameters and minimized prediction deviations between imported and produced loads, this article proposes first, a method for forecasting based on a regression ensemble method: adaboost regressor and second, a comparison study between hybrid approaches based on LSTM, MLP and the model proposed with adaboost regressor. This last study provides the basis for an optimal selection of methods for future forecasting of electrical consumption loads. So, this article is divided into two parts: the first consists of the learning and validation tests of the proposed model, and the second is a comparative study between the proposed model based on adaboost and those, based on Convolutional Neural Networks (CNN), MLP, Long Short-Term Memory -Attention Mechanism (LSTM-AM) and LSTM-AM-MLP. The data used in this article were collected from a renewable energy source: photovoltaic solar energy. While 80% of the data collected was used for learning purposes, the remaining 20% was used for validation testing. The results of the first part of this study give a coefficient of determination R² between 0.9995 and 0.9997 for the learning results and between 0.919 and 0.958 for the validation test results. These results are representative of the real data and reflect the performance of the proposed model. In the second part of this study, the prediction results obtained by the comparison of the proposed model with another one using LSTM-AM-MLP, show a prediction gap ranging from 0.15 to 0.78 for RMSE and 0.83 to 2.07 for MAPE. These results show the significant differences between the methods. The error minimization of the proposed model reflects its accuracy. The proposed model is well adapted to the management of electrical consumption load forecasts to ensure balance between supply and demand.