I implementeda GRU-based Seq2Seq model with attention to predict three quantiles of residential power demand in Texas using historical residential electricity consumption (source: ERCOT Hourly Load Data Archives) and 10-year weather data (source: Weather Underground). The model achieved high accuracy (6% MAPE, 0.35 MAE), outperforming existing literature.
The forecasts were subsequently used as inputs for stochastic optimization models, supporting energy planning and trading applications.
📍Github Repository: https://github.com/FarzaneEzzati/PD-Seq2SeqRNNAttention
📍Github Repository: https://github.com/FarzaneEzzati/PD-Seq2SeqRNNAttention
[Picture]
Quantile Loss (three quantiles 0.1, 0.5, 0.9) for power demand in Austin city over training and validation.
[Picture]
Quantile Loss (three quantiles 0.1, 0.5, 0.9) for power demand in Houston city over training and validation.
[Picture]
MAPE (mean absolute percentage error) in % for prediction in Houston city over training and validation.
[Picture]
MAE (mean absolute error) of prediction in Houston city over training and validation.