- β
- Width scaling factor; each network shrinks by β−φ.
- Band
- The 10% error budget above each baseline, the most accuracy an energy saving was allowed to cost.
Formulation
Each morning, the day-ahead market clears a price for every hour of tomorrow. The model reads the past 24 hours of price, load, weather and time features and returns all 24 next-day prices at once, as one vector, so errors do not compound hour by hour. Training uses the Huber loss, which behaves like squared error on ordinary hours and like absolute error on the rare price spikes.
Data
Four years of hourly ERCOT load and day-ahead prices for the Houston zone, joined with NOAA air temperature and wind speed. Hour, weekday and month enter as sine and cosine pairs, so 23:00 sits next to 00:00. The last 10% of the series, about four months, is held out and scored in sliding windows.
Models
The team compared an RNN, an LSTM, a CNN-LSTM, an Informer-style Transformer, SARIMA and a SARIMA model with LightGBM fitted to its residuals. The CNN-LSTM puts two causal one-dimensional convolutions in front of stacked LSTMs, which picks up sharp local changes the LSTM alone smooths over. A second input carries tomorrow's time encodings into the decoder.
Compound scaling
To cut energy, the networks shrink the way EfficientNet scales them, with depth multiplied by and width by :
Energy per epoch was measured with Zeus on Colab T4 GPUs, the same hardware anyone can rent.
Results
The CNN-LSTM had the lowest error of the deep models, 18.07 $/MWh RMSE against the LSTM's 20.10, and caught the size and timing of evening price spikes that the LSTM overshot or smoothed. It also degraded gracefully when shrunk: at β = 1.1 it used 4% less energy within the 10% error budget, and at β = 1.9 it used 12% less, at a cost in accuracy. The LSTM's error nearly doubled at the first step.