Multistep-Ahead Solar Irradiance Forecasting for Smart Cities Based on LSTM, Bi-LSTM, and GRU Neural Networks

Abstract

Sustainable and renewable energy sources provide a promising method to address worldwide energy crises due to their long-lasting availability and a clean environment. However, this solution has drawbacks in optimizing energy production and demand integration. For instance, the intermittent nature of photovoltaic system power influenced by weather conditions is the most significant obstacle to appropriate integration into smart city systems; hence, among these sustainable resources, solar irradiance requires accurate prediction. Therefore, this study proposes recurrent neural network (RNN)-based deep learning models for time-series forecasting problems to reflect nonlinear weather parameters effectively. These methods include long short-term memory (LSTM), bidirectional LSTM (Bi-LSTM), and gated recurrent units (GRU), along with their variants for 11-step-ahead (one day) hourly solar irradiance forecasting for Seoul, Busan, and Incheon. The performance of these methods was evaluated by comparing them with baseline regression models comprising multiple linear regression, partial least squares, and multivariate adaptive regression splines based on the mean absolute and root mean square errors. In addition, the variants of RNNs were compared in terms of performance indices. Attention mechanism-based Bi-LSTM and GRU models trained with the scaled exponential linear unit activation function derive excellent performance in multistep-ahead solar irradiance forecasting. A comparison with the existing results supports the proposed RNN variants due to their higher efficiency, accuracy, and robustness.