A CDW-LSTM Based Multi-Step Prediction Method for Superconducting Magnet Temperature

The superconducting magnets are widely used, and the temperature is an important critical condition. The refrigeration process is very important, necessitating copious amounts of liquid nitrogen and helium as cooling agents. Accurately multi-steps predicting the temperature during the refrigeration process is crucial for ensuring the stability of the magnet's state and enhancing refrigeration efficiency. Furthermore, due to the potential for multiple adjustments in both hardware and software parameters, rapid adjustments to the model and the generation of suitable parameters are equally significant. LSTM is widely used to predict the time sequence of temperature and it is shown outstanding performance. In order to improve accuracy of long-term temperature prediction, a Cauchy mutation and double adaptive weight optimized whale algorithm (CDW) to adjust the hyperparameters of the LSTM algorithm is proposed and validated. With the quick convergence speed and overall search capability for parameter optimization, CDW is used to adaptively adjust the hyperparameters of the Long Short-Term Memory (LSTM) network, and combined with a direct-recursive prediction strategy, it further improves the accuracy of multi-step temperature prediction.