Learning Compact Neural Network for Long-Term Time Series Forecasting

Time series forecasting plays an important role in many science and engineering fields, such as electricity consumption and weather conditions prediction. Recently, linear models have shown great performance in long-term time series forecasting, which even surpasses transfomer and its variants. However, the number of parameters increases linearly with the length of both the input and output time series, thus bringing about overfitting problem to the model and incurring difficulties for handling long sequences on resource constrained devices in terms of storage and computational consumption. In this paper, we propose a Learning scheme to automatically discovery Compact model for Long-term time series Forecasting, which is dubbed as LCLF. Firstly, a compact network architecture with bottleneck structure is constructed for time series forecasting. Then L1-regularization is introduced to learn the optimal dimensionality of the bottleneck layer during model training. Lastly, a compact and sparse model can be obtained without perceptible perrformance degradation. Extensive experiments are carried out on six long-term time series tasks, and the results demonstrate that the proposed method achieves better performance comparing to other competitive approaches.