Solar Wind Speed Prediction Based On Multi-Step And Multi-Task Learning

Solar wind is an important part of space weather.The prediction of solar wind speed is of great significance to people’s production and life.With the rapid development of machine learning and artificial intelligence in recent years,deep learning methods are more and more used in the prediction of solar wind speed.At present,the mainstream deep learning solar wind speed prediction technology usually adopts the single-step prediction to predict the solar wind speed at a certain time point in the future according to the historical data.In the field of multivariate time series prediction,multi-step prediction is a widely used method with many advantages.Aiming at the problem of deep learning solar wind speed prediction,the following work is carried out in this paper:Firstly,a traditional multi-output multi-step prediction method is adopted.In view of the characteristics that the recent local historical data have a great impact on the 24-hour short-term prediction,two multi-step prediction models are constructed: single LSTM model and encoder-decoder LSTM model,which can capture the local timing characteristics for prediction.In view of the characteristics that the long-term global historical data has a great impact on the 96 hour long-term prediction,a multi-step CNN-LSTM model is constructed to better capture the global spatial characteristics for prediction.We conducted experiments using a 6-year solar multivariate time series data set.The results show that the multi-step prediction of each model is better than its single-step prediction,which verifies the effectiveness of multi-step prediction.Then,we use the multi-task learning method as an optimized way to realize multistep prediction,and construct the multi-task solar wind speed prediction model ”MMSWSP”,which shares the shallow features of multi-step output and constructs the deep features separately for prediction.The experimental results of MMSWSP were compared with the above traditional multi-step prediction and single-step prediction models,and the experiments of selecting the length of input window,as well as three groups of ablation experiments using only single task,removing corresponding modules and using only part of input parameters were carried out.The results show that the prediction accuracy of MMSWSP is better than the above models,which verifies the effectiveness of multi-task learning.In conclusion,combined with the characteristics of solar wind speed prediction,this paper constructs effective multi-step and multi-task deep learning models.Among them,the multi-task model has achieved the effects of CC of 0.843,RMSE of 59.881 km/s and MAE of 43.362 km/s in short-term 24-hour prediction,and CC of 0.668,RMSE of 88.682km/s and MAE of 68.783 km/s in long-term 96-hour prediction.