Load Analysis And Prediction Based On MapReduce And Deep Learning

Intelligent power utilization, an important constituent in overall architecture of smart grid, aims to provide more personalized and humanized intelligent service through the two-way interaction between power supply enterprises and users. Accurate power load analysis and forecast is the premise of intelligent power service. However, with the laying of smart meters, smart terminals and all kinds of sensors, the data in the user side has reached the scale and complexity of big data, greatly enhancing the difficulty of load analysis and forecasting. How to effectively mining potential knowledge and pattern from these massive data, thus providing a reliable basis for decision making when power supply enterprises try to expand power market, optimize energy structure, and rationally allocate resources, is currently one of the most challenging task in power system.Big data technology and deep learning theory is the key to solve the above problems. This paper introduced a universal framework of intelligent power big data platform, based on the analysis of the relationships among the big data, deep learning and intelligent power utilization. The key elements of the platform were explored in four aspects, including the data source layer, data processing layer, mining analysis layer and business application layer. Then, considering the traditional decision tree algorithm cannot solve the massive data, a parallel decision tree classification algorithm based on the MapReduce framework was applied to mine the habits based on the smart meter data. Also, the correlation coefficients between the attributes were introduced to avoid the multi-value bias problem of ID3 algorithm. The effectiveness of the algorithm was verified through the actual load data according to the mining results and parallel performance.Deep learning can fully excavate the law of the massive load data and achieve a more accurate short-term load forecast. This paper introduces the stacked autoencoder(SAE) and deep belief network(DBN) from three aspects:basic components, network structure and the training methods. An algorithm based on information entropy was proposed to determine the number of the hidden units, and the effectiveness of this method was validated through the MNIST data set. Also, based on the distributed memory computing platform Spark, these deep neural networks were trained with multiple model replicas. Finally, the typical load curves were used to improve the performance of load forecast, and the feasibility of the overall method is verified through the load data by comparing with other algorithms.