(Deep) Learning Energy Price Volatility

Volatility in wholesale electricity prices presents risk to utility firms constrained by local regulators to keep retail prices within approved bounds. Knowledge about future price fluctuations might help firms mitigate that risk. We use a deep learning neural network, an architecture of convolutional and long short-term memory neurons (CNNs, LSTMs), to predict day-ahead wholesale price volatility over two years of hourly data from PJM, a U.S. regional transmission organization. These data, in conjunction with price volatility time series generated by a GARCH (generalized autoregressive conditional heteroskedasticity) process, were used to train a deep learning model composed of CNN and LSTM layers. Segregated testing data were then used to evaluate the trained model and measure its generalizability. Compared to a naive baseline method and a simple multi-layer perceptron model with one standard fully-connected hidden layer, the deep learning model outperformed both in terms of mean squared error (MSE) for delay targets 3, 6, 12, and 24 hours into the future. This suggests it may be possible for utility firms to anticipate short-term price volatility through the development of similar but proprietary models (wherein additional input features such as weather data and natural resource prices might also be included) to hedge against demand shocks and competitor behavior