|
Abstract
|
Short-term load forecasting (STLF) enhances the power system's economic efficiency and operational stability. Due to their limited ability to handle multi-scale features and complex temporal dependencies, existing energy forecasting models often fail to effectively capture local and global patterns in time-series data. There is a need for a more robust forecasting approach that integrates multi-scale feature extraction and advanced temporal modeling to improve scalability and accuracy in energy forecasting. This paper proposes a novel hybrid model of Convolutional Neural Networks and Gated Recurrent Units (CNNGRU) model for improved electric energy forecasting, which integrates multi-scale Conv1D layers to extract spatiotemporal features at different scales and GRU layers to model temporal dependencies. Therefore, This approach is useful in modeling both local and global patterns and can be considered a good solution to the problem of time series analysis in energy forecasting. APE and ISONE publicly available datasets are used to carried out a thorough analysis to evaluate the effectiveness of the proposed model. Proposed model findings show a consistency across all key matrices in comparison with other models, achieving the lowest Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) values in both single-step and multi-step. With the AEP dataset, the model produced single-step forecasting results of RMSE of 131.91, MAE of 77.39, and MAPE of 0.67, and multi-step forecasting results of RMSE of 686.32, MAE of 486.17, and MAPE of 3.24. Likewise, about the ISONE dataset, the model achieved a single-step forecasting RMSE of 184.29, a multi-step forecasting RMSE of 588.03, a multi-step forecasting MAE of 297.94, and a MAPE of 0.52. This demonstrates the superior accuracy and robustness of the proposed model across different prediction horizons.
|