Probabilistic grid carbon intensity forecasting with Hodrick–Prescott decomposition

Reducing grid carbon emissions plays a crucial role in addressing climate change and achieving sustainable energy development. Accurate forecasting of grid carbon intensity contributes to adjusting energy consumption strategies and enhancing clean energy utilization. Due to the uncertainty of renewable energy output, carbon intensity often exhibits non-stationarity and stochasticity. To this end, we propose a probabilistic carbon intensity forecasting method based on time series decomposition. First, a Hodrick–Prescott filter is studied to extract different components from the original carbon profile. Second, a long short-term memory (LSTM) with pinball loss is leveraged to produce quantile forecasts to the cycle component with greater volatility. Probabilistic forecasting results can be obtained by ensembling the forecasts of all components. Extensive experiments conducted on two open datasets demonstrate the superior performance of the proposed method.