Central Air-conditioning Load Forecasting And Energy-saving Optimization Control Based On Deep Reinforcement Learning

With the vigorously coordinated development of China’s economy and other aspects,China’s energy consumption is a major problem that needs to be addressed,building energy consumption accounts for a relatively high proportion among it,and the proportion is still rising.Therefore,Research on energy-saving optimization of buildings has received extensive attention in recent years.Among all types of building energy consumption,central air-conditioning consumes a relatively high proportion of energy,so effective energy-saving optimization of central air-conditioning is a major measure to reduce the energy consumption of buildings in our country.In order to further improve the control strategy of central air-conditioning,this discertation uses the Deep Deterministic Policy Gradient(DDPG)algorithm in deep learning to obtain a better control strategy,so as to achieve effective energy saving.However,the deep deterministic policy gradient algorithm has the problems of unstable dual network structure and inaccurate evaluation by a single critic,which will make the final effect of practical application unsatisfactory.This discertation makes a series of improvements to the original DDPG algorithm,and proposes a DDPG algorithm based on multiple exponential moving average evaluations.At the same time,it uses the Fuzzy C-means algorithm(Fuzzy C-means,FCM)to optimize the Double Deep Q Network(Double Deep Q Network,DDQN)and predicts the load of the air-conditioning,and provides optimal control through the predicted load value at the next moment,so as to achieve significant energy saving.It mainly includes the following three parts:(1)For the problem of the instability of the dual network structure and the inaccurate evaluation of a single critic in the DDPG algorithm,a DDPG algorithm based on multiple exponential moving average evaluation is researched.The algorithm obtains the target update value through the cooperation of the EMA-Q network(Exponential Moving AverageQ Network)and the target Q network,and averages the Q values given by multiple critics to reduce the inaccuracy of single critic evaluation.In addition,the sample pool part introduces a double experience replay method,two sample pools are used to store different experiences separately to improve the convergence performance of the algorithm.In the experiment,the DDPG algorithm based on multiple exponential moving average evaluation is applied to inverted pendulum and mountain car,and compared with the original DDPG algorithm.(2)Aiming at the air-conditioning load with time sequence and high interference properties,a load forecasting method based on FCM-optimized DDQN is researched.Compared with the original load forecasting model for the entire sample,the FCM algorithm first clusters the samples,and then the load forecasting model is constructed according to the different types of samples.The samples after FCM algorithm clustering are input into the forecasting model to be more representative,and the subsequent load forecasting accuracy is also higher.In order to verify the effectiveness of the forecasting method,the real-time operation of air-conditioning was simulated by inputting the actual building air-conditioning load data recorded by a certain environmental school into the forecasting model,the load forecasting method of optimized DDQN after FCM clustering and not introduced FCM optimized DDQN load forecasting method experimental data are used for comparative analysis.(3)Combining the DDPG algorithm based on multiple exponential moving average evaluation with the predicted next interval load,the air-conditioning system can be better controlled and optimized.First,a brief analysis of the central air-conditioning system is made.The DDPG algorithm based on multiple exponential moving average evaluation is used to optimize the variables for different load intervals.The variables selected are the chilled water outlet temperature,the cooling water flow rate,and the chilled water outlet and return temperature difference,and get the optimal parameter value.Predict the specific value of the load at the next moment in advance according to the FCM-DDQN algorithm,and select the appropriate action variable to realize the energy-saving optimization of the airconditioning system.