Conv-AdaRNN:A Power Load Forecasting Method Based on CNN and AdaRNN

• Published in: 2022 5th International Conference on Hot Information-Centric Networking (HotICN) pp. 72 - 76
• Main Authors: Zihan, Wang, Enze, Shao, Can, Wang
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24.11.2022
• Subjects: CNN; Feature extraction; Load forecasting; power load forecasting; Predictive models; RNN; Temperature; Time series analysis; time series data; Transfer learning; Wireless communication
• ISSN: 2831-4395
• DOI: 10.1109/HotICN57539.2022.10036174

Accurate power load prediction is crucial for power dispatch and response. In order to improve the accuracy of power load prediction, this paper proposes a new method called Conv-AdaRNN which combines Convolutional Neural Network and Adaptive Learning and Forecasting for Time Series(AdaRNN) models. The method firstly uses CNN to extract valid information from effective factors, that affect the power load, such as temperature, humidity, and electricity price to construct feature vectors. Then, the data set is divided into the least relevant segments in time series according to the Time Series Distribution law. Finally, the AdaRNN, whose hidden states are assigned different weight to match time distribution, analyzes feature vectors belong to each segment and forecasts power load. Taking a certain area's power load as an example, the predicting results of the Conv-AdaRNN model were compared with those of gated recurrent unit (GRU) and AdaRNN models. The results show that, compared with other single prediction models, the Conv-AdaRNN achieved the better forecast precision.