| dc.description.abstract | In a real-time scenario of load forecasting, it is crucial to determine the future electric energy consumption in power distribution electrical networks. The electric energy forecasting models need to be updated with real-time trends of energy consumption as the analyzed energy consumption data exhibits high variability between historical and current data. This work proposes a multi-stage supercomputing-based big data analytics service for parallel and real-time load forecasting. Moreover, theoretical and experimental perspectives are proposed for multi-core parallel short-term load forecasting. Additionally, the knowledge from existing load forecasting based on deep learning models is used to innovatively develop highly accurate transfer learning models at different distribution nodes. Transfer learning models present practical applicability and productive possibilities in cases when sufficiently large data is not available. A novel approach based on deep neural network models is employed for load forecasting. Firstly, the electrical distribution nodes are grouped into different clusters with an aim to decrease the number of deep learning models to be trained. Secondly, network architecture information, weights, and biases are transferred from the first developed clustered model to subsequent models with an aim to reduce the training time of a large number of clustered models. And incremental learning is employed to incorporate newer data points for real-time processing and improving the forecasting accuracy of the clustered models on individual distribution points. Furthermore, parallel pool-based processing is employed to make efficient utilization of computing cores and to reduce the model development time further. The proposed big data real-time analytics methodology is evaluated on real-world energy consumption data collected from 105,148 Spanish electrical distribution transformers. The proposed methodology aims to reduce the number of trained models, training time, and execution time while still maintaining high prediction accuracy. | |
