Solar Irradiance Forecasting Using Runge-Kutta Based Continuous-Time Recurrent Neural Network

Abstract

Solar Energy (SE) is an efficient renewable (RE) energy source owing to its environmental benefits. Global horizontal irradiance (GHI) is an important component for SE-based applications. Therefore, this energy source is widely employed to fulfill electricity demand through stable power generation in everyday grid operations. This growing adoption plays a major role in the power supply market. To ensure an uninterrupted power supply from this sustainable power source, accurate and reliable GHI forecasting is necessary. However, the non-stationary and intermittent nature of the GHI makes traditional statistical and machine learning (ML)-based models less effective at providing accurate forecasts. Recently, advanced forecasting architectures such as RNNs, LSTMs, GRUs, Bi-LSTMs, and Bi-GRUs have been shown to perform well in GHI forecasting. However, these models yield suboptimal forecasting results with respect to error metrics. Because these models primarily rely on extracting features from time-series data, they do not capture the continuous dynamical changes in historical data, resulting in lower forecasting accuracy. To address the aforementioned limitation, this study proposes a Runge-Kutta-based Continuous-Time Recurrent Neural Network (CTRNN) architecture that learns the data’s dynamical behaviour more effectively and yields better results. The effectiveness of this architecture has been explained by applying CTRNN to a multivariate SE dataset. This study evaluates its performance and compares it with the aforementioned models using error measures such as MAE, MSE, RMSE, MAPE, and R2-score. The corresponding error metric values of the proposed model are MAE: 0.02945, MSE: 0.00336, RMSE: 0.05796, MAPE: 1.14056, and R2-score: 0.97854. The results of GHI forecasting using CTRNN indicate that the proposed model outperforms other models.