Forecasting solar energy production in Spain: A comparison of univariate and multivariate models at the national level

Renewable energies, such as solar power, offer a clean and cost-effective energy source. However, their integration into national electricity grids poses challenges due to their dependence on climate and geography. While numerous studies have focused on solar energy time series, few have specifically addressed the critical task of forecasting solar energy production at the national level. Accurate national-level forecasting is crucial for optimizing energy management, informing policy development, and promoting environmental sustainability. This study aims to address the challenges associated with the significant variability in renewable energy production and its impact on grid stability by improving the accuracy of existing forecasting approaches. To achieve this goal, we evaluate the effectiveness of univariate and multivariate approaches for time series forecasting of national solar energy production data from ESIOS (the Spanish System Operator). Our primary focus is on leveraging external solar variables, such as solar irradiance data. To this end, we propose a methodology to integrate solar irradiance forecasts with historical data from solar power plants in Spain to improve the performance of multivariate models. Subsequently, we compare the performance of classical regression techniques and state-of-the-art deep learning algorithms, presenting univariate and multivariate models for three forecast horizons (1 h, 24 h, and 48 h). Finally, we assess the performance of our best univariate and multivariate models by comparing them with the official forecast of ESIOS. Our findings indicate that the best-performing models are deep-learning multivariate approaches, which benefit from incorporating solar irradiance forecasts, particularly for longer forecast horizons (24 h and 48 h), and avoid the detrimental effects of the Hughes Phenomenon, which seems to hamper non-deep-learning forecasters. The top-performing models, based on Convolutional Networks and Convolutional + Recurrent Neural Networks, outperform ESIOS by reducing mean absolute error by 41% and 47.58%, respectively.