A Temperature Time Series Forecasting Model Based on DeepAR

Abstract

In order to analyze the massive data and complex non-linear relationship in temperature forecasting, we apply deep learning techniques to forecasting. In this paper, by viewing the original weather data as numerical time series, we propose an improved network model based on DeepAR to produce probabilistic forecasts. We use stacked temporal convolution layers and LSTM layers in an encoder-decoder architecture to accelerate the training and predicting process. Experiments on a real-world weather dataset shows that the proposed model has a good improvement on accuracy compared to DeepAR and common-used baselines.