Energy Reports (Nov 2022)
Towards efficient and effective renewable energy prediction via deep learning
Abstract
Renewable energy (RE) offers major environmental and economic benefits compared to nuclear and fuel-based energy; however, the data used for RE include significant randomness, intermittent behaviour, and strong-volatility, hindering their integration into smart grids. Accurate RE prediction is a promising solution to this problem and can provide effective planning and management services. Various predictive models have been developed to improve the prediction performance for better energy management. However, current works focus on improving the prediction accuracy, which is a requirement of power systems, without considering the time complexity of their methodologies. Considering these limitations, we develop a lightweight ESNCNN model for accurate RE prediction, in which an ESN learns the nonlinear mapping relationship and a CNN extracts the spatial information from RE data. The ESN and CNN layers are linearly connected via residual connections to avoid the vanishing gradient problem. Finally, we incorporate fully connected layers to enhance and select the optimal ESNCNN features to predict the future energy production. Our ESNCNN model is evaluated based on RE benchmarks, using various evaluation metrics such as MSE, MBE, MAE and RMSE, and achieves state-of-the-art performance. Further experiments are performed with different machine learning, deep learning, and hybrid models to select the optimum model. To fully assess the generalisation ability of the proposed ESNCNN, additional experiments are performed over electricity consumption datasets, which reveal an extensive decrease in error rates compared to other state-of-the-art approaches. Our model therefore represents a new paradigm for finding an energy equilibrium between grid and energy production resources using a single ESNCNN platform. Our results indicate a substantial reduction in the error rates over the RE dataset (MSE (5.01%), MAE (5.49%), and RMSE (3.76%)) and the electricity consumption dataset (MSE (5.37%), MAE (7.63%), RMSE (0.047%), and MBE (1.2%)).
