A Study On Heating Load Forecast Based On Artificial Neural Network And Support Vector Machine And Their Modified Algorithms

Centralized heating was put force in the area where need heating inwinter is the main technique for the development of China’s Urbanization;however the heating energy consumption in the proportion of social energyconsumption is much large. With the increasing emphasis on energyconservation, how to improve the heating efficiency has become the primaryproblem to reduce the consumption of heating energy.To be more accurate to predict the heating load is the key to improvethe heating efficiency, so that the heating enterprises can produce and distributeheating energy according to needs. The research of heating load forecast hasmade considerable progress and development after introducing the modernartificial intelligence prediction technology, but it still need to be further studied,such as using which algorithm and how to improve them.. Two intelligentprediction algorithms, the most widely used artificial neural network and thenew prediction technology of support vector machine in the most promising,was put forward to the research of heating load forecast. Different optimization prediction model was set up, combined with the development trend predictiontechnology and method, to observe and analyze the effect of its prediction forheating load forecast.For the artificial neural network algorithm, the BP neural networkprediction model and its improved model, wavelet neural network, were used toforecast heating load by studying the wavelet analysis theory. At the same time,for the support vector machine algorithm, the support vector regressionprediction model was used to forecast heating load. After analyzing thecontrolling parameters effect on the predicted results, the cross validation whichis a method for digging through the affordable data, and the grid search, geneticalgorithm and particle swarm optimization, which were three method forsearching best parameters, were put forward to establish the model ofGS-KCV-SVR, GS-KCV-SVR and PSO-KCV-SVR for heating load forecast.The training set and the test set for all the prediction model used thesame heating field data. Considering all the relevant factors for the input of themodel, theoretical analysis was used for the influence factors of the outdoorenvironment, and qualitative analysis was used for the internal factors of thesystem related to the heating load, and quantitative calculation was used toverify the analysis results. Considering users active impact on heating loadchange after joining the metered heating, the day type parameter was introducedas one of the input parameters of the heating load forecast model. The moresuitable input and output function expression of heating load forecast model was established combined with the intelligent prediction algorithm of highdimensional computational advantages and the heating system of complexthermodynamic characteristics.The analysis and calculation results of various heating load forecastmodels showed: the principle of support vector machine is more advanced thanthe artificial neural network in the high dimension problem of multi factorsrelated to heating load; the grid search is less efficient but easy to realize;genetic algorithm and particle swarm algorithm in optimization of performancehave their own merits, so selecting the best appropriate optimization algorithmneed depending on the specific situation; the prediction accuracy of theoptimized forecast model is better than the original prediction model; theprediction accuracy used the model of support vector regression and itsoptimization algorithm is better than that of used the model of neural networkand its whole optimization algorithm; the algorithm of cross validation is helpfulto improve the convenience and promotion of prediction model. In terms of thestability of the model, support vector regression model is higher than BP neuralnetwork model which needs repeated calculation to get satisfactory result, andwith the affected by random factors, genetic algorithm and particle swarmalgorithm is less than the grid search method.The various prediction models were built in the research of heatingload forecast, which would provide effective reference to scientific productionand the necessary basis for the heat source distribution and scheduling for heating enterprises. Through the analysis and comparison of each model，basedon the measured data of the heating and the various aspects of the evaluationfactors, it recommended: the heating load forecast model of GS-KCV-SVR,which performs balanced in stability, convenience and accuracy, was used topredict the heating load with the general scale of sample data; the heating loadforecast model of GA-KCV-SVR or PSO-KCV-SVR was used to predict theheating load with the large scale of sample data and some special requirements.