The Research On Key Techonology And Application Of Time Series Forecasting Based On Machine Learning

With the rapid development economy and the acceleration of globalization,largescale,multi-category,and multi-dimensional massive data have been generated in various fields such as industrial production,economic activities,climate change,energy production and distribution,and transportation.Among them,time series data objectively reflect the changing process of different phenomena and activities in various fields,and there are rich variation laws behind it.We can learn and summarize related data variation patterns and laws by analyzing and studying these data.If we apply these laws to production and other activities,we could realize multiple purposes,such as energy-saving and consumption reduction,risk avoidance,and optimization of decision-making.Therefore,time series forecasting plays a vital role in modern society and is of great practical significance and practical value.In addition,with the rapid development of science and technology,we are about to enter the Io T Age,which will inevitably bring more types and larger volumes of time series data.Time series forecasting will face a large number of data samples,diversity of data types,and multiple forecasting targets,which makes time series forecasting very challenging research.Since the variation of time series data has specific periodic laws and,at the same time,is also affected by various random factors,the fluctuations of them are often characterized by nonlinear and random,and have a complex periodicity,which makes accurate time series forecasting becomes ever more difficult.However,when traditional time series forecasting methods face the challenges of large volume,multiple categories,and rapid changes of data brought about by the dramatic growth of modern time series data,it is increasingly challenging to meet application requirements.Time series forecasting methods need to be further improved.To improve the forecasting effect and application value of time series forecasting methods,we consider various fields' practical applications and conduct in-depth research on some critical issues in time series forecasting.The main research contents and innovative work are as follows:(1)Breaking through the limitations of traditional time series forecasting methods in selecting feature data sets,a time series forecasting method based on source-trace data is proposed.To explore the internal mechanism and causes of the fluctuations of data,we collect features from multiple dimensions.Then,feature selection is performed from a large number of features that have been confirmed to be related or maybe potentially related to the fluctuations of data to construct a source-trace feature data,which provides comprehensive and three-dimensional data support for the prediction model.Case studies based on daily peak load data from Maine and Texas of the USA,and electricity price data from New South Wales of Australia demonstrate that the proposed method could obtain better forecasting results.(2)We also propose a time series forecasting method based on cross multi-model and second decision mechanism.First,aiming at solving the problem that traditional continuous training data set cannot reflect the data changes in the long term and that the similar day-based discrete training data set cannot effectively capture the trend of data in the short term,we propose a cross training set construction method that constructs the training sets from both horizontal and longitudinal directions.Based on cross training sets,the second learning method is applied,and the second decision mechanism is proposed by applying multiple decision models.The second decision mechanism could effectively increase the generalization ability and extend the application scope of the proposed method.Case studies are conducted based on the daily peak load data of Singapore,the half-hourly load data of New South Wales of Australia,the daily total load of New England region of the USA and the traffic flow data of California,and the accuracy and stability of the proposed method are demonstrated.(3)To improve the integration accuracy of multi-model methods,we propose an adaptive weight allocation strategy to perform weighted integration on multi-models' outputs.The weights are calculated according to the prediction accuracy of different independent prediction models on the validation set,and the extreme values of the model output results are eliminated before the integration.Compared with the traditional average integration strategy,performance-based integration strategy,and competitive optimization integration strategy,this method has the advantages of being insensitive to outliers and high integration accuracy.A case study based on the daily peak load of Maine demonstrates that the proposed adaptive weight allocation strategy could provide accurate,stable,and robust integration results for multi-model methods.(4)A time series forecasting method based on source-traced data and cross multimodel is finally proposed by combining all the innovative work in feature data set construction,training set selection,model construction,and multi-model integration.At the feature data set construction level,the source-traced feature dataset is constructed to systematically explore the source factors that cause time series data variation and provide rich information for the forecasting model.At the training set level,the cross-training set reflects the fluctuation law of time series data in different directions and provides diverse sample information.At the model construction level,the application of the second decision mechanism effectively improves the model's performance in the decision stage and the generalization ability of the method and expands its application scope.At the multi-model integration level,the adaptive weight allocation strategy is applied to assign higher weights to independent models with better performance and improve the accuracy and stability of the results.