Ventilation Scheduling Strategy Research Based On Deep Reinforcement Learning

In recent years,with the rapid development of cities,the air quality in cities is getting worse and worse,and people are paying more and more attention to indoor air quality.Most of the modern new office buildings are equipped with ventilation systems,that is,the outdoor fresh air is filtered and purified and sent indoors,it is significant for ventilation system to provide efficient and energy-saving scheduling decisions for different ventilation scenarios.However,the indoor ventilation scene is greatly affected by personnel activities and outdoor air quality,and the decision-making itself also requires extremely high effectiveness.It has become an urgent problem to construct an energy-saving and efficient ventilation scheduling strategy and conduct formal verification and analysis on it.In view of the above problems,this paper proposes a ventilation scheduling strategy based on deep and reinforcement learning,and conducts a verification analysis on it.First,the data change regular of the room is obtained by learning the historical data of the room through the LSTM+Attention neural network,and the K-Means clustering and TF-IDF algorithm are used to analyze the meaning of the data to generate the ventilation user portrait;then,analyze the user portrait and use Q-Learning algorithm generates a response ratio scheduling strategy,a priority-based local scheduling strategy and a global scheduling strategy,and uses the UPPAAL-SMC tool to model and verify the analysis of the scheduling strategy.Finally,through the specific case of the building ventilation scheduling scenario,user comfort and energy consumption of different scheduling strategies are evaluated,and the applicable scenarios of each scheduling strategy are summarized.The main work of this paper is as follows:·In order to better delineate target users and understand user demands,this paper proposes to analyze the overall ventilation usage of the room by constructing ventilation user portraits.First,the LSTM+Attention neural network is used to learn and predict the historical data of indoor air quality parameters,such as CO2,TVOC,PM2.5,and HCHO,to obtain the data changes of different rooms at various times.Secondly,K-Means clustering and TF-IDF algorithm are used to determine ventilation scenarios represented by different data,to judge the overall indoor air quality and distribution of people.·In order to cope with the changes of ventilation in different scenarios,this paper proposes three ventilation scheduling strategies based on the ventilation user portraits of each scenario,which are the response ratio scheduling strategy based on reinforcement learning,the local scheduling strategy based on priority and the global scheduling strategy.The core strategy is the response ratio scheduling strategy based on reinforcement learning,which uses the Q-Learning algorithm to dynamically update the response ratio coefficient according to environmental changes,which has the most extensive application scenarios;the priority-based local scheduling strategy adheres to the principle of maximizing energy,the population density of the room is scheduled in descending order;the global scheduling strategy is that when the overall air quality is judged to be poor,the room will turn on the fresh air indiscriminately.·This paper use UPPAAL-SMC to model and analyze the fresh air dispatching strategy.In order to compare the effectiveness of different dispatching strategies in practical applications,this paper divides the ventilation scenarios into six categories according to the outdoor air quality and user classification.The user discomfort probability,user cumulative comfort time and ventilation energy consumption are compared and evaluated,and the best applicable scenarios of different scheduling strategies are summarized.