Aiming at the problem that the existing electricity load forecasting methods fail to fully capture electricity load data's periodicity and trend characteristics

resulting in poor prediction stability

this paper proposes a short-term electricity load forecasting model PowerTimeMixer. Firstly

the original time series is decomposed

and the periodicity and trend characteristics of the time series are learned in a multi-scale manner; secondly

the convolution down sampling mechanism is introduced to match the cycle period through network parameter sharing

further enhancing the feature extraction ability of the periodic pattern of electricity load data; finally

an independent date pattern driven forecasting module is introduced

and the multi-layer perceptron is used to encode the timestamp features of the input sequence and the target sequence. The timestamp features are learned independently

guiding the network to generate more stable prediction results according to the timestamp features. The experimental results on the electricity load dataset show that the prediction error of the proposed method is significantly reduced compared with the baseline model

and it has a more stable prediction performance

which verifies the effectiveness of the proposed method.