Research On Greenhouse Microclimate Prediction Model Based On Improved BP Neural Network

It is the basis and key technology to establish an accurate prediction model of greenhouse microclimate for optimizing the control of the greenhouse environment.This paper adopts Auto-Encoder(AE)and the improved particle swarm algorithm(IPSO)to optimize the BP neural network model.The research object are greenhouse in the Middle East and South of Jiangsu.This study constructs a real-time rolling prediction model and medium-long term prediction model of greenhouse microclimate based on the improved BP neural network.The validity of the model is verified by comparing with traditional methods.The main contents of this paper are as follows:(1)Key factors influencing the greenhouse microclimate are determined by correlation analysis.We choose outdoor environmental factors,indoor equipment state,and indoor temperature and humidity at previous moment as the network input,and then construct a single step prediction model of greenhouse microclimate in the Middle East and South of Jiangsu.The greenhouse microclimate real-time rolling prediction model and medium-long term prediction model are established based on the model.(2)Aiming at the defect that BP neural network is easy to fall into local optimum,an AE+IPSO_BP single step prediction model is proposed in this paper.Firstly,the initial network parameters are obtained by AE unsupervised learning,and then IPSO algorithm is used to optimize neural network weights and threshold parameters.The IPSO algorithm takes advantage of local version particle swarm optimization algorithm with multi particle swarm to search in the solution space independently,and combines with ideology of genetic algorithm(GA).The crossover operation is introduced to the particle position during the process of updating algorithm,and the mutation operator is introduced to the particle whose fitness value is lower than the average value of the population in many times,reducing the possibility that the PSO algorithm falls into local optimum.(3)In order to solve the problem of cumulative error in the traditional medium-long term forecasting methods,this paper proposes a medium-long term prediction model of the greenhouse microclimate based on R_BP neural network.According to the prediction time,the R-BP model constructs a BP neural network every 15 minutes,and then forms a rolling BP neural network group.The model includes two stages.The first stage is to establish an initial BP neural network model(AE+IPSO_BP model is adopt),and get better prediction results as well as network parameters.In the second stage,based on the previous network model,several single step prediction models on the basis of BP neural network are established.The output of previous network is used as partial input of the next network to carry on the rolling training and prediction,to achieve the medium-long term forecast of greenhouse microclimate.(4)In the real-time rolling prediction test,we compare the AE+IPSO_BP model with traditional BP neural network model,each model running 10 times,and the final result is the average root mean square error(RMSE)of 10 times.In the test of Abu Dhabi greenhouse,AE+IPSO_BP is compared with the BP neural network,with 14%decrease of average RMSE on temperature and 13.5%decrease of average RMSE on relative humidity.In the test of Suzhou greenhouse,after the comparison of AE+IPSO_BP and BP neural network,the average RMSE of temperature reduces by 37.5%and the average RMSE of relative humidity reduces by 18.4%.The experimental results show that the AE+IPSO_BP model can accurately predict the short-term change of greenhouse microclimate,so as to realize the real-time control of greenhouse microclimate,and to lay a good foundation for the medium-long term prediction.In the medium-long term prediction test,we compare the R-BP model with traditional BP neural network model,each model running 10 times,and the final result is the mean value of 10 prediction errors.In the test of Abu Dhabi greenhouse,we use R-BP model to predict the temperature and humidity inside of the greenhouse in the next 6 hours.Meanwhile,R-BP is compared with the traditional BP neural network,with 69.9%decrease of average prediction error on temperature and 47%decrease of average prediction error on relative humidity.In the test of Suzhou greenhouse,after the comparison of R-BP and traditional BP neural network,the average prediction error of temperature reduces by 43.3%in the next 6 hours and the average prediction error of relative humidity reduces by 55.6%.The experimental results illustrate that the R-BP model is practical and effective.It can accurately predict the medium-long term greenhouse environmental changing in different seasons and different regions.Furthermore,the prediction results provide the basis for establishing reasonable microclimate regulation scheme.