Leveraging Temporal Patterns in Electricity Price Forecasting with Cluster-Based Feature Engineering and Temporal Fusion Transformers

This paper introduces a novel methodology for electricity price forecasting by leveraging an adaptive clustering method with attention-based learning in real-time electricity markets. Generation clusters are identified from all power generators in an energy market using the Hierarchical Density-Based Spatial Clustering of Applications with Noise to separate these clusters. These clusters, along with the market's price, supply, and demand variables, are incorporated as features into a Temporal Fusion Transformer model to generate forecasts up to 6 hours ahead. When applied to forecast electricity prices in Ontario and compared to the system operator's forecasts, the proposed model reduces the mean average error and root mean square error by 46% and 30%, respectively, on average on the entire forecast window. The performance during volatile periods is also increased, shown in the improvement of precision above 71% during first 3 hours while maintaining recall above 56%, both when prices above the percentile 90% during the test period are forecasted. Furthermore, the developed methodology can be applied to other markets and adapt to structural changes such as the expansion and modernization of generation fleets.