Pulse-diagnosis-inspired multi-feature extraction deep network for short-term electricity load forecasting

Forecasting short-term electricity load (STEL) is a very important but challenging task by the fact that the series dynamic change involves in multiple patterns, such as long short-term cyclical, nonlinear abrupt, nonlinear gradual, and linear monotonic patterns, while interactions of the above patterns are unknown and variable. In addition, most existing deep forecasting models only focus on the data-layer fitting, and inherit unexplained flowchart and domain knowledge lacking. Motivated by that traditional Chinese medicine (TCM) practitioner analyzes pulse signal coupled with multiple patterns, i.e., pulse strength, thickness, length, rate, etc. and further diagnoses diseases at the extraction-feature level, this paper develops a pulse-diagnosis-inspired multi-feature extraction deep network with three following components. The sense block mimics that the TCM practitioner applies fingers to feel and extract multiple pulse features, the comprehension block mimics that the TCM practitioner integrates the above pulse features and hence grasps core abnormal features, and the judgement block mimics that the TCM practitioner deduces patient's diseases based on core abnormal features. Specifically, the first block is applied to intelligently extract multiple pattern features based on the domain knowledge, the second block is used to fuse the above features and obtain core features via the gate mechanism, and the third block is adopted to make final forecasts for STEL. The proposed network learns the pulse diagnosis in both flowchart (layer by layer) and function (three block) aspects, producing explained flowchart and clear block function. Experimental results under two real-world electricity load series with different sampling intervals (30 and 60 min) show that the proposed deep network has higher accuracy and stability than 12 baselines, and improves the root mean absolute percent error by averages of 42.9% and 55.8%, respectively. Thus, the proposed deep network provides reliable electricity evidence for power sector scheduling.