Application of a hybrid ARIMA-LSTM model based on the LightGBM in grid frequency prediction

Accurate grid frequency prediction is crucial for maintaining power systems' stability and reliability. This paper proposes a hybrid ARIMA-LSTM model based on a Light Gradient Boosting Machine (LightGBM) to predict the time-series grid frequency, which can produce more reliable prediction results and enable timely measures to reduce grid frequency anomalies. Ultimately, this will ensure better, safer, and more stable power grid operation. First, a method based on the combination of exponential moving average (EMA) and discrete wavelet transform (DWT) is used to reduce the influence of noise on the data so that the wavelet transform can decompose the data more effectively. Subsequently, an Autoregressive Moving Average Model (ARIMA) and a Long and Short Term Memory Network (LSTM) are used to predict the trend and period portions of the decomposed data. Finally, LightGBM combines the prediction results from the ARIMA and LSTM models to achieve comprehensive time series prediction results with high accuracy and reliability. The proposed method's effectiveness and accuracy are verified using a real grid frequency dataset. The experimental results show that merging the prediction results using the LightGBM model can produce satisfactory prediction performance, which is superior to other prediction methods. In conclusion, the application of this method significantly improves the accuracy of grid frequency time series prediction.