Study On Prediction Model Based On Hadoop For Building Energy Consumption

Building energy consumption, industrial energy consumption and transportation energy consumption are the three major energy consumption of the society. Building energy consumption has been a great burden that restricts our economic development. Predicting the energy consumption of exsiting buildings can be used to analyze and modify a designing scheme to achieve a low-energy design which has a guiding significance for building energy efficiency design. In order to find some useful information from the large scale building energy consumption data we need some data analysis techniques and data mining techniques. Traditional data prediction algorithms focus on small samples. If the building information has a large scale, the calculation would increase sharply and the training time would lengthen, which are problems that traditional algorithms cannot be overcome. Because of the raise of cloud compting, we may process some traditional prediction algorithms in parallel to solve the large-scale data prediction effectively. So the main content in this thesis as follows:(1) The building energy consumption data we have collected has some characteristics like large-scale, non-uniform, irregular, high- dimensional and so on. According to the characteristics, we use rough set theory to reduce attributes of data sample to find the main factors affecting the building energy consumption. Compared to the serial reduction algorithm, the parallel reduction algorithm based on Map Reduce can achieve the same reduction results and shorten the data processing time.(2) According to the characteristics of the sample we have pre-processed, we fix the topology of BP neural network predicting the building energy consumption, such as the number of layers, the number of neurons per layer, transfer function, learning algorithm and so on, to build the prediction model of building energy consumption. The common BP neural network algorithm has some disadvantages like longer training time, easy to converge to local minimum and so on. Using genetic algorithm to optimize the building energy consumption prediction model, and making comparisons with the common model. We can find that the improved model is better than the common model in the prediction precision and the training speed from the experiment. The model is only suitable for small samples. When the data size is large, it will be hard to train.(3) According to the model based on RBF, we conduct an experiment with the input and output of this thesis and find that the speed of RBF is faster than BP, but the accuracy did not improve greately. Compared to the BP algorithm optimized by GA, the speed and accuracy of RBF did not show some advantages. So we chose BP algorithm to construct the prediction model for building energy consumption and optimize the initial weights and thresholds by GA.(4) Studying the parallel points of common BP neural network to make an extension based on Map Reduce to build a prediction model focusing on large-scale building information and deploy it on Hadoop distributed platform. Applying the BP neural network algorithm based on Map Reduce proposed in this paper for the large building energy consumption data prediction and the initial weight and threshould in the experiment choose the results optimized by Genetic Algorithm. On the premise that the prediction accuracy can be ensured, compared with the traditional BP neural network, the algorithm in the dissertation has an unparalleled advantage in large-scale building information prediction. Also the operation rate of the new algorithm is better than the traditional algorithm.The innovation of this thesis is mainly reflected in two aspects:(1) Using a data reduction algorithm based on Map Reduce to preprocess the large-scale building energy consumption data.(2) Building a parallel BP neural network model based on Map Reduce for large-scale building energy consumption data.