Long Short‐Term Memory Network and Statistical Time Series Analysis Forecast Models for 30 min Interval Wind Farm Power Output and Regional Price Variables

This study compares parametric statistical time series models, such as autoregressive moving average (ARMA), with nonparametric artificial neural networks, specifically long short‐term memory (LSTM) models, for univariate forecasting. Two time series are analyzed separately: wind power output from the Clements Gap wind farm and the regional electricity price for South Australia. One‐step‐ahead forecast performance is evaluated using normalized mean bias error (NMBE), normalized mean absolute error (NMAE), and normalized root mean square error (NRMSE). Three LSTM models were examined: a manually tuned model, a structurally equivalent model implemented in a different library, and a model with automated hyperparameter tuning. While LSTM models achieved competitive performance, statistical models often performed equally well or better. For price forecasts, the manually tuned LSTM achieved the lowest NMBE (7×10−6$$ 7\times {10}^{-6} $$), while ARMA(3,1) had the best NMAE (0.0253) and AR(6) the best NRMSE (0.220). For wind forecasts, the manually tuned LSTM again performed best overall (NMBE: 0.00034, NMAE: 0.143, NRMSE: 0.225), while the equivalent library model performed worst. These results highlight the need to subject nonparametric LSTM models to more rigorous and systematic evaluation relative to their parametric statistical counterparts.