CNN Assisted Dual Reservoir Hybrid Network for Power Consumption Forecasting

Developing a precise forecasting model for power usage is an essential requirement for modern power management systems. As a result, researchers are focusing their efforts on proposing effective and accurate energy forecasting techniques which are crucial for future energy management and planning in smart grids. The prevailing methods for energy forecasting rely on traditional methods which are challenged by the high level of nonlinearity among input data from smart sensors and output, requiring further improvements in terms of robustness, forecasting performance, and generalization ability. Therefore, in this study, a novel hybrid approach combining CNN and a dual reservoir is developed for future energy consumption prediction. Unlike existing methods, our model leverages CNN to extract spatial dependencies in historical data, which are then processed by a dual reservoir to capture complex temporal relationships. This unique combination enhances robustness, forecasting accuracy, and generalization ability. Additionally, Bayesian Optimization Methods (BOM) are employed to fine-tune hyperparameters, further improving performance. The proposed model is systematically developed through extensive experimentation with both standalone and hybrid models, demonstrating superior predictive capability compared to traditional approaches. The proposed model achieved RMSE values of 0.0245 and 0.0332 for hourly building data, and 0.0113 for regional data, which are significantly lower than the baseline methods, demonstrating its superior performance.