There are various deep learning methods already implemented in time-series forecasting problems

and some of them show better performance and adaptability than the methods based on statistics. However

the spatial dependence implicit in multiple series usually is not considered within the existing time-series forecasting methods

which also results in unsatisfactory performance in long-term forecasting. Thus

we propose a model series-decomposition-block recurrent neural network (SDBRNN)

which fuses time-series decomposition strategy and the spatio-temporal convolutional layer. This model relies on an improved STL decomposition strategy to get optimal series-components and to combine them into spatio-temporal blocks. Then an long-short term memory (LSTM) variant network called Block-LSTM is used to extract spatio-temporal features from blocks and achieve forecasting. Experiments on real-world datasets proved that the model has excellent capabilities of feature extraction and long-term forecasting compared with other methods such as temporal convolutional network and recurrent neural network.