Applying Probabilistic Forecasting of Multidimensional Time Series to Electricity Load Analysis

Electricity load is a crucial time series for electricity power systems, and its analysis and forecasting are essential for making informed decisions regarding future power infrastructure construction and power generation. Current deterministic point forecasting methods for electricity load forecasting have limitations. Probabilistic forecasting, on the other hand, provides more comprehensive analysis information by offering the interval and probability distribution of future changes. However, existing probabilistic schemes for time series forecasting, such as Variational Auto-Encoder (VAE), suffer from time-consuming training, large volume of model parameters, and inability to capture long-term dependencies in series. To address this challenge, this paper proposes a new probabilistic forecasting model V-trans that combines the self-attention mechanism and VAE. The model uses the self-attention scheme to capture long dependencies and VAE to extrapolate probability distribution in prediction results while minimizing computation costs and model storage. Evaluation experiments demonstrate that the proposed model has higher forecasting accuracy.