A short-term power load prediction method based on Attention-LSTM in the context of optical-computing synergy
TL;DR
The Attention-LSTM model has performed better than the Persistence, LSTM and GRU models in terms of MAE, RMSE and MAPE, and according to the ablation experiment, the attention layer and some kinds of modal input variables are also useful.
Abstract
Accurate short-term forecasts of the electrical power demand are required for the intelligent operation of the power grid and energy management. With the introduction of more decentralized power units and fluctuating energy demand, power consumption now has considerable non-linearity and time dependency, so the old prediction method is no longer suitable. To improve the accuracy of the forecast and operational efficiency, a new short-term energy consumption prediction method using Attention-LSTM is put forward in this paper. An optical interconnect system is employed in the above structure to achieve high-speed, stable transmission of consumption data; data cleaning and abnormal detection are done at the edge, and model and hyperparameter optimisation are carried out in the cloud. The dataset of daily energy demand is used for the test of the algorithm. The chosen input factors are past consumption, weather conditions and time attributes; a rolling window set is created, and then the demand for the next hour is predicted. LSTM is used to find the time dependency of the power profile, and attention is added to highlight important historical information. According to the simulation results, the Attention-LSTM model has performed better than the Persistence, LSTM and GRU models in terms of MAE, RMSE and MAPE. According to the ablation experiment, the attention layer and some kinds of modal input variables are also useful. Therefore, there will be more precise and punctual support for demand forecasting in the operation of intelligent grids in the environment of optical-computing synergy.
