Prediction Of The Amount Of Refrigerant Remaining Based On Long Short-term Memory

With the rapid development of productivity,energy consumption is also increasing.Buildings consume about 20~40% of the world’s energy consumption,with heating,ventilation and air conditioning accounting for about 40% of building energy.VRF(variable refrigerant flow)system,as an efficient cooling system that minimizes operating and maintenance costs,combined with their high flexibility,comfort,and reliability,are now widely used in buildings.However,the aging of equipment and various faults will have a greater impact on its performance.Due to the complexity of VRF system,there is a variety of equipment,it is impossible to diagnose all kinds of faults.Refrigerant leakage is a common problem in VRF,previous studies have focused on the diagnosis of refrigerant charge,and there were few studies on the prediction of refrigerant content.Therefore,this thesis uses a data-driven method to predict the amount of refrigerant remaining in VRF system.Since the content of refrigerant is difficult to measure during system operation,this thesis uses the empirical relationship established by the predecessors to calculate the refrigerant content at the current time based on the actual operating data.In order to obtain steady-state data,Gaussian steady-state discrimination algorithm is used to extract steady-state data.The results show that the proposed method can eliminate the unsteady data in the original data.By using Pearson correlation coefficient and Maximal information coefficient to select 5 features with high correlation,as the input characteristics of LSTM(long short-term memory)model and BP(back propagation neural network)model,a VRF system refrigerant residual prediction model is established.The results show that after the original data is cleaned by the steady-state discriminant algorithm,the trend fluctuation caused by the refrigerant leakage is well preserved,which improves the quality of the data and is beneficial to improve the prediction accuracy of the model.Before the model optimization,the MAE(mean absolute error)and MSE(mean square error)of the LSTM model are smaller than the BP model,and the prediction curve is more consistent with the actual trend,but the LSTM model is larger in the calculation cost.After using the dynamic learning rate and adding the Dropout layer,the performance of both models is improved.The MAE of the LSTM model is increased from 0.93 to 0.38,and the BP model is increased from 2.89 to 1.12,indicating the validity of dynamic learning rate and Dropout layer.In order to be closer to the application,multi-step prediction is performed using the LSTM model to predict data for 10-time steps after the current time.At the same time,in order to test the generalization ability of the LSTM model,the model is validated using data under different working conditions,and finally the difference between the effects of univariate and multivariate time series prediction is compared.The VRF refrigerant residual prediction model proposed in this thesis introduces the deep learning algorithm LSTM into the field of refrigeration and air conditioning and achieves better prediction results,which can provide reference for the maintenance personnel of the unit to develop maintenance plans.It can be applied to other VRF systems in the future and it is worth exploring.