Attention enhanced dual stream network with advanced feature selection for power forecasting

Energy forecasting is crucial for balancing electricity demand and supply, enabling efficient management and effective planning in the smart grid and other power management systems. Several hybrid networks with diverse applications in the smart grid have been deployed; nonetheless, their forecasting efficacy is limited due to a lack of optimal data refinement, feature extraction, and selection. Therefore, this paper presents an intelligent framework a Dual-Stream Deep Network (DSDN), for power generation and consumption forecasting, which is mainly composed of two phases. The initial phase employs various preprocessing methods to remove outliers, impute missing values, and apply data normalization to minimize data deviation. The second phase involves a parallel integration of the Echo State Network (ESN) with a Self-Attention Module (SelfAM) and a Residual Convolutional Neural Network (RCNN) enhanced by a Spatial Attention Module (SpatialAM). The DSDN parallel structure facilitates the extraction of spatial and temporal features from actual historical data while the skip connections helps to mitigate the vanishing gradient issue, and the attention modules enable the network to capture salient features across both dimensions. Subsequently, the outputs of both streams are then concatenated into a unified feature vector, which is processed through Principal Component Analysis (PCA) for optimal feature selection and dimensionality reduction, followed by fully connected layers for final forecasting. The performance of DSDN is assessed using various evaluation metrics on benchmarks for power generation and consumption. The results reveal that the proposed model achieved the highest performance compared to baseline methods for both generation and consumption forecasting. The DSDN exhibited the error value of 0.08929 MAE and 0.14209 RMSE over the DKASC dataset, while 0.01958, 0.033, and − 0.00985 for MAE, RMSE and MBE, respectively, over the IHEPC dataset. Furthermore, DSDN is not solely evaluated on a specific dataset, but on a combination of multiple datasets, including photovoltaic, residential, and industrial power consumption. The higher performance of the DSDN across these diverse datasets underscores its versatility and efficacy, making it a robust solution for a wide array of smart grid applications.