Convolutional-LSTM networks and generalization in forecasting of household photovoltaic generation

Solar panels can generate energy to meet almost all of the energy needs of a house. Batteries store energy generated during daylight hours for future use. Also, it may be possible to sell extra electricity back to distribution companies. However, the efficiency of photovoltaic systems varies according to several factors, such as the solar exposition at ground levels, atmospheric temperature, and relative humidity, and predicting the energy generated by such a system is not easy. This work is on the use of deep learning to predict the generation of photovoltaic energy by residential systems. We use real-world data to evaluate the performance of LSTM, Convolutional, and hybrid Convolutional-LSTM networks in predicting photovoltaic power generation at different forecasting horizons. We also assess the generalizability of the solutions, evaluating the use of models trained with data aggregated by geographic areas to predict the energy generation by individual systems. We compare the performance of deep networks with Prophet in terms of MAE, RMSE, and NRMSE, and in most cases, Convolutional and Convolutional-LSTM networks achieve the best results. Using models trained with region-based data to predict the power generation of individual systems is confirmed to be a promising approach.