The economic impact of the innovative forecast of solar energy production based on machine learning and neural network models

Solar power, characterized by intermittency and weather dependence, is becoming a significant component of the renewable energy mix. Carrying out a systematic review of solar energy estimation strategies, this study aims to analyse the economic impact of innovative models designed to enhance the accuracy of solar energy forecasts using advanced Machine Learning and Neural Network models for two photovoltaic systems. The results show that the advanced forecasting models outperform a persistence benchmark based on the same hour of the previous day. Concerning Machine Learning models, Light Gradient Boosting Machine and Categorical Boosting highlight consistent performance, indicating greater robustness compared to traditional models such as Decision Trees and Random Forest due to their ability to reduce bias and variance. When comparing Neural Network with Machine Learning models, Long Short-Term Memory shows higher robustness (reducing RMSE by more than 35%). Furthermore, the study introduces clustered models for weather-based data groups and demonstrates superior performance (RMSE 20.87). This method significantly reduces forecast errors compared to using a single Machine Learning model (RMSE XGBoost model for “Clear” 43, GBM model for “Partly cloudy” 39.41, CatBoost model for “Overcast” 32.22). The economic analysis reveals that grid operators are able to reduce the costs associated with grid balancing and the interventions required to maintain grid stability by reducing forecast errors. The monthly quantitative deviation decreased for both photovoltaic systems (32.34% to 18.07% and 36.44% to 22.41%). Accurate models facilitate the optimisation of energy storage capacities and resource utilization, leading to operational efficiency, and cost control.