Research On Image Sequence Prediction And It's Hyperparameter Optimization Method

In recent years,the rapid development of Internet technology has continuously improved the quality of people's life,which has led to higher requirements for the application of new technologies.High-precision weather forecasting,automatic driving and other tasks have become urgent problems to be solved.With the rapid development of machine learning,especially deep learning algorithms in computer vision,image prediction technology evolves with the time.In the current field of meteorological forecasting,radar echo extrapolation is a widely used technology in short-term rainfall forecasting services.In practice,the radar echo extrapolation is commonly achieved by traditional algorithm,such as the optical flow method,which has a poor long-term extrapolation effect and a low utilization rate of information in the radar echo data.In this paper,a deep learning method is proposed,together with a model based on pyramid generative adversarial network(P-GAN).The network model adds multi-scale network and multiple loss function cooperative training,it also establishes the mapping relations between the input image sequence and prediction result images.Experiments were carried out on the S-band doppler radar dataset collected in Guangdong Province in the past nine years.The experimental results show that,compared with the traditional radar echo extrapolation method represented by optical flow method,the prediction model designed in this paper can obtain higher predictive image accuracy.Multiple indicators have been improved significantly and the model is capable of obtaining more realistic forecast images.Due to the complexity of the problem,the proposed image sequence prediction model is time-consuming to train,and the number of hyper-parameters is large.In order to manually optimize the hyperparameters,it requires a lot of repeated attempts.Furthermore,some popular algorithms in the machine learning field,such as the grid search,random search and Bayesian optimization algorithm,also requires a lot of time to iteratively find a better combination of hyperparameters.In order to solve this problem,this paper improves the Reverse-Mode Automatic Differentiation(RMAD)and proposes a real-time hyperparameter automatic optimization algorithm.By continuously calculating the gradient of the loss function to the target hyperparameter in the model training process,the algorithm automatically optimizes the hyperparameter based on the gradient,and only needs one longer round of training process to complete the optimization for the target hyperparameter.This algorithm reduces the time required to repeatedly adjust the iterative adjustment of the hyperparameters,and it is proved by experiments that the method can improve the prediction effect of the model.