Multi-Task Time Series Forecasting of Plant Load Losses: A Comparative Study of a Temporal Fusion Transformer and LSTM
TL;DR
Evaluating the TFT relative to an LSTM to forecast load losses 16 weeks ahead and predict the plant that will likely fail, across three data scales reveals the empirical boundaries of transformer-based models, illustrating that data volume serves as a pivotal determinant for the effective activation of advanced self-attention mechanisms of the TFT.
Abstract
Accurate load loss forecasting is important to prevent plant failures and improve power station reliability. While the Temporal Fusion Transformer and LSTM have demonstrated state-of-the-art performance in modelling complex temporal patterns, their effectiveness remains strongly influenced by the scale of the available data, and despite their promise, their application to predict load losses in power stations is underexplored, particularly for medium-term forecast horizons. We evaluated the TFT relative to an LSTM to forecast load losses 16 weeks ahead and predict the plant that will likely fail, across three data scales: 2000, 10,000, and 50,000 derived from a 5-year secondary load loss dataset from six power stations. The TFT substantially outperformed the LSTM on all data sizes, with the 50,000-sample size achieving the best results of 0.9789 prediction accuracy, 12.3853 MSE, 0.5539 MAE, 3.5193 RMSE, and 0.9858 for R2. Collectively, these findings uncover the empirical boundaries of transformer-based models, illustrating that data volume serves as a pivotal determinant for the effective activation of advanced self-attention mechanisms of the TFT.