Seasonal Scale Climate Encodings For Spatio-temporal Demand Forecasting

Abstract

Incorporating seasonal climate insights in time series forecasting problems, such as demand predictions, can help inform planning and optimizing operations. Current time series forecasting approaches incorporate deterministic short-term weather attributes as exogenous inputs. However, encoding the relationship between seasonal climate and demand is challenging due to the uncertain nature of seasonal predictions and their associated spatio-temporal variability and predictive skills. Recently, time series research has introduced a deep learning-based temporal fusion transformer (TFT) model using self-attention for modelling different types of time series. In this work, we incorporate seasonal climate predictions in TFT and experimentally observe that forecast errors can be reduced by 5-17% on real-world dataset while forecasting up to few months ahead.