The Weakly Usable Multivariate Time Series Prediction Method Based On Scenario Clustering And EMD

With the era of Big Data coming, the forecast has become the core of Big Data. However, with a large of date is collecting, the weakly usable data, which can’t be corrected completely, such as noisy data, missing data and ambiguous data, is brought into the multivariate time series, and constitutes the weakly usable multivariate time series (WUMTS) which makes the prediction results inaccurate and the results of data mining wrong finally. Therefore, how to predict the WUMTS effectively, accurately and robustly is the key and core researched issue in this master thesis. The research topic has important academic value and practical significance in the big data era.Based on analyzing the characteristics of WUMTS, the thesis proposes a prediction method suited for WUMTS from three aspects of WUMTS preprocessing, combination forecasting, and error correcting according to the fragmentary, inexactitude characteristics of WUMTS. The main contents are as follows:(1) According to the problem of common prediction method never mind weakly usable data, the thesis proposes a WUMTS prediction method. The framework of the WUMTS prediction method is based on the traditional prediction method framework and it consists of the WUMTS preprocessing, the fine-grained decomposing, combination forecasting and the component prediction results recomposing.(2) According to the fragmentary and inexactitude characteristics of WUMTS, the thesis proposes a WUMTS preprocessing method. It is realizing to integrate, reduce and transform the WUMTS, fill the missing data, denoise base on sparse decomposition and reduce dimensionality base on principal component analysis method. It constructs the WUMTS preprocessing module and improves the usability of WUMTS partly.(3) According to the combination forecasting need to decompose the WUMTS effectively, the thesis proposes a WUMTS fine-grained decomposition method included the multivariate time series decomposition based on scenario clustering with factors and the multivariate time series decomposition based on EMD with series values. It reduces the sensitivity of factors and correlation of the series effectively in WUMTS to construct the WUMTS fine-grained decomposition module.(4) According to the adaptability problem between prediction algorithm and WUMTS, the thesis proposes a combination forecasting method based on Adaboost algorithm. Through adaptively selecting the alternative weak predictor and combining them to construct a strong predictor with Adaboost algorithm. The method realizes the combination forecasting of the WUMTS fine-grained decomposition component, and constructs the combination forecasting module based on Adaboost algorithm finally.(5) According to the cumulative error problem of EMD linear additive recomposing, the thesis proposes a nonlinear component prediction result recomposing method based on SVM optimized parameters. Through the MPCSO optimizes the SVM’s recomposing parameters, the method realizes the nonlinear recomposing with component prediction results which can correct error partly and constructs the component prediction results composing module based on MPCSO-SVM.Finally, through the cooling load WUMTS experiment, every module has been verified to the superiority, and does not produce the cumulative error among the modules. Therefore, the feasibility, accuracy, robustness and universality of the WUMTS prediction method proposed in the master thesis has been verified.