Deep Learning-Based Hybrid Model for Forecasting Locational Marginal Prices

It has generally become a challenging task to accurately forecast locational marginal prices (LMPs) due to its strong volatility originated from transmission congestion and extreme weather. Most of existing literature fails to predict both price spikes and normal prices with equally satisfying effectiveness. In fact, normal price prediction mostly relies on the strong correlation with demand, which proves to have a high Pearson correlation coefficient with load changes. However, price spikes prediction is associated with network congestion resulting from features with high Maximal Information Coefficient (MIC). In this paper, we propose a deep learning-based hybrid model for LMPs prediction, considering the discrepancy of two kinds of price. Based on a support vector machine (SVM) and a given threshold, we firstly distinguish and label normal prices and price spikes. For normal prices, a long-short term memory (LSTM) neural network can uncover the trend of time series and bring out an accurate prediction. For price spikes, a light gradient boosting machine (LightGBM) is adopted to address those extreme jumps by analyzing relevant features in tabular format. Case studies based on realistic datasets from the PJM market demonstrate the efficiency and effectiveness of our proposed hybrid model.