Training method of bidirectional LSTM (long short term memory) model for implementing locomotive energy-efficient operation

Abstract

The invention provides a training method of a bidirectional LSTM (long short term memory) model for implementing locomotive energy-efficient operation; the method comprises the steps of acquiring driver historical driving data and locomotive operation monitoring log as initial training data; preprocessing the initial training data to obtain a training dataset and a test dataset; in case of initial training, initializing parameters of a model directly; otherwise, continuing to train a bidirectional LSTM neural network model based on the previous bidirectional LSTM model, and saving the trained model; using the trained model to perform simulation testing on the test dataset to obtain new initial training data; using the new initial training data and the training data of the previous training phase jointly as initial training data; repeating the steps until the model converges. The feature design method, model design method and iterative training method provided herein allow data information to be fully utilized, and gear prediction capacity of the model is improved.

Description

technical field [0001] The invention relates to the field of energy-saving manipulation of railway locomotives, in particular to a bidirectional LSTM model training method for realizing energy-saving manipulation of locomotives. Background technique [0002] The locomotive operation control system is a typical multi-objective, multi-constraint, nonlinear complex control system, which needs to ensure reliability, safety, punctuality and fuel economy, so the problem of locomotive energy-saving maneuvering is a nonlinear and constrained multi-objective Dynamic optimization problems. However, since many complex constraints need to be considered in the solution process of this type of problem, the entire optimization search space is very large, and it is difficult to search for the optimal solution in a short period of time, and its research has great practical significance. [0003] The existing optimization methods for locomotive energy-saving maneuvering operation can be divi...

AI Technical Summary

Problems solved by technology

Among them, the numerical search refers to the optimal search of the manipulation sequence through the numerical search algorithm to obtain the optimized manipulation sequence. Common algorithms include genetic algorithm, group search algorithm, dynamic programming, etc., but this method takes a long time and is difficult to converge to the optimal The optimal result; the analytical solution method refers to solving the key transition points in different situations in the manipulation control process based on domain knowledge according to the analytical formula to obtain the final optimized manipulation sequence. It is difficult to deal with multiple constraints; the heuristic strategy design method refers to the consideration of many complex factors, and manual heuristic design through some operating specifications in the existing field. This method introduces too much manual analysis and design, which is extremely difficult. The efficiency of strategy design is greatly reduced, and at the same time, due to the limited scope of human thinking, it is impossible to cover all possible situations, which will inevitably lead to the omission of some optimal solutions.

Images