BOL-LPP: A Bayesian-Optimized LSTM Model for Day-Ahead Load Price Forecasting in the ERCOT Market

Abstract

Precise short-term load price forecasting is critical for uninterrupted and efficient powersystem operation and energymarket performance. Although machinelearning techniques have been widely employed to predict market prices, achieving reliable dayahead load price forecasts remains challenging in practice, especially in the Electric Reliability Council of Texas (ERCOT) energyonly market. This article targets sufficiently accurate dayahead load price prediction for ERCOT’s zonal markets by modeling historical load, price, and weather data with a Long ShortTerm Memory (LSTM) network whose hyperparameters are tuned via Bayesian Optimization (BO). The resulting BayesianOptimized LSTM for load price Prediction (BOLLPP) is evaluated against classical statistical and deeplearning baselines. On the Northzone test set, BOLLPP attains a Mean Absolute Error (MAE) of ${\$}$0.0044/MWh, cutting the MAE by 32% relative to the strongest deep baseline (BiLSTM, MAE of ${\$}$0.0065/MWh) and by over 99% compared with SARIMAX. Its MAE remains below ${\$}$0.006/MWh on the Coast and South zones, confirming robust generalization. These numerical results, along with the reported Mean Squared Error (MSE), and Mean Absolute Percentage Error (MAPE), validate the performance gains delivered by the proposed model. BOLLPP therefore promises markedly improved shortterm load price forecasts, supporting informed decisionmaking and enhanced operational efficiency in the ERCOT market.