A Multi-Task End-to-End Multivariate Long-Sequence Time Series Prediction Model for Load Forecasting

With the increasing complexity of the power system, and the growing global demand for electricity, accurate and effective forecasting of electricity load, electricity prices, and related meteorological features becomes increasingly crucial. However, existing methods tend to use only one element, for example, electricity load, as their prediction target instead of generating multiple indexes (e.g., electricity load, electricity prices, and related meteorological features) predictions jointly. This is insufficient for electricity producers to make decisions. While some time series forecasting methods can predict multiple indicators simultaneously, they often overlook the correlations among these features. These approaches miss the potential performance improvement that modeling correlations among multiple series could bring to the prediction results. To address the aforementioned concerns, this study presents a model named PatchGRU. Once trained, PatchGRU is capable of predicting long-sequences of electricity loads, electricity prices, and related meteorological features in a single forward propagation. The model differentiates between time-variant and time-invariant components using a Temporal Variance Separator, processing them separately. For the time-variant part, the network uses the proposed multi-scale patch input in place of traditional point inputs, then feeds it through the Gate Recurrent Unit (GRU), mean supervision, and Local-Global Feature Interaction modules, ultimately generating the forecast results. Moreover, this model relies solely on GRU and Multilayer Perceptrons (MLPs) and is much simpler than the Transformer structure. Experimental results show that the proposed model outperforms state-of-the-art models in different regions, time periods, and sampling frequencies, achieving the best results in forecasts with a maximum lead time of up to 30 days.