Solar energy forecasting using liquid time-constant networks

This paper investigates the application of Liquid Time-Constant (LTC) networks for solar energy forecasting. Traditional models, such as ARIMA and conventional machine learning techniques, often fail to effectively capture nonlinear relationships and process continuous-time data. We implement LTCs, a continuous-time recurrent neural network, as a novel solution and benchmark their performance against other deep learning models, namely LSTM, GRU and a vanilla ANN. Utilizing a dataset comprising weather variables and solar energy readings, we assess model performance using R-Square, Mean Square Error (MSE), and Mean Absolute Error (MAE). The LTC model excels with an R-Square of 0.9820 and MSE of 0.5264, surpassing other models in these metrics, while the LSTM model achieves a slightly better MAE of 0.5543 compared to LTCN's 0.5937. These results highlight the LTC's superior ability to model complex temporal patterns, positioning them as a promising tool for solar energy forecasting with enhanced accuracy over existing approaches.