Abstract:

In order to improve the prediction accuracy of air temperature in the solar greenhouse, Genetic algorithm (GA) was introduced to optimize the initial weights and thresholds of BP neural network. Using hourly meteorological forecast data (air temperature, relative humidity, and wind speed) outside the greenhouse and air temperature data of 1−3 hours before the forecast inside the greenhouse, the hourly prediction models of the temperature in the solar greenhouse with earth and brick walls were conducted based on GA−BP neural network. Simulation results by GA−BP neural network were compared with that simulated by the stepwise regression (SR) and BP neural network models. The results showed that in the greenhouse with earth wall, the root mean square error (RMSE) and the normalized root mean square error (NRMSE) between observed and simulated hourly air temperature with the SR model were 0.82−2.01℃ and 5.13%−9.97%, respectively. The RMSE and NRMSE between observed and simulated hourly air temperature with the BP neural network model were 0.82−1.79℃ and 5.13%−7.74%, respectively. The RMSE and NRMSE between hourly air temperature observed and simulated hourly air temperature with the GA−BP neural network model were 0.62−1.47℃ and 3.88%−6.40%, respectively. In the greenhouse with brick wall, the RMSE and NRMSE between observed and simulated hourly air temperature with the SR model were 1.07−2.60℃ and 5.11%−16.98%, respectively. The RMSE and NRMSE between observed and simulated hourly air temperature with the BP neural network model were 1.11−2.29℃ and 6.12%−14.95%, respectively. The RMSE and NRMSE between observed and simulated hourly air temperature with the GA−BP neural network model were 0.89−1.73℃ and 4.76%−11.30%, respectively. The error index values representing the difference in observed and simulated hourly air temperatures with the GA−BP neural network model were lower than that with the SR and BP neural network models. The introduction of genetic algorithm to optimize BP neural network can improve the prediction accuracy of hourly temperature. The temperature fluctuation in the greenhouse with the brick wall was greater than that in the greenhouse with the earth wall because the thermal insulation of brick wall was slightly worse than that of earth wall. Therefore, the temperature forecast accuracy for the greenhouse with earth wall was higher than that for the greenhouse with brick wall. The forecast error increased with the advance of the leading time of forecast. The GA−BP model established by selecting hourly air temperature, relative humidity, and wind speed outside the greenhouse and air temperature data of 1−3 hours before the forecast inside the greenhouse as modeling factors had the best accuracy. GA−BP neural network model could be used to predict hourly temperature inside the solar greenhouse with high accuracy and stability.

Key words: Genetic algorithm, BP neural network, Stepwise regression, Air temperature, Solar greenhouse, Dynamic forecast