Study On Mechanism And Methods Of Adaptive Streamflow Forecasting Based On Machine Learning

Under the impact of changing environment,the formation process and evolution mechanism of streamflow are complex and changeable,the stationarity assumption is no longer valid.Based on historical data and surveys of actual changes,the traditional work of evaluating consistency,reliability,and representativeness has exposed m any problems.In this context,it is difficult and critical to carry out streamflow forecasting that responds to changes.Driven by these problems,taking the advanced and feasible technological route is both an exploration and a breakthrough to streamflow forecasting.Based on the adaptive streamflow forecasting in response to environmental changes and supporting decision-making,this thesis has carried out research on streamflow forecasting methods and adaptive forecasting mechanisms.Based on the big data analytics technologies,this thesis firstly analyzed the evolution law of streamflow,mined the driving factors of streamflow formation,and extracted predictors for streamflow forecasting.With the aid of machine learning,this thesis secondly proposed a model construction method for adaptive streamflow forecasting and established streamflow forecast models based on statistic law and cause law of streamflow.Supported by the integrated support platform for knowledge visualization,this thesis finally designed and implemented an adaptive streamflow forecasting system,which realized the adaptive streamflow forecasting under the integration of different application themes,different driving factors,different spatial and temporal scales,and dif ferent forecasting models.Focusing on the adaptive streamflow forecasting service,the following main works has been carried out:1.An adaptive streamflow forecasting mechanism has been designed.Aiming at the problem that the stationarity assumption is no longer valid under the changing environment,the impact and demand of the changing environment and water resources management on streamflow forecasting were first analyzed.The change objects and factors that need to be responded by adaptive streamflow forecasting are then clarified.An adaptive streamflow forecasting mechanism,which provides new patterns and ideas for streamflow forecasting in changing environment,was designed from three aspects of streamflow forecasting,i.e.,the input variables,the model,and the application of prediction results.A series of technical methods to realize the adaptive streamflow forecasting was constructed based on big data analytics and machine learning.A knowledge-map-based integration realization technology of adaptive streamflow forecasting was established based on integrated support platform for knowledge visualization.These methods laid a technical foundation for adaptive streamflow forecasting.2.A method for mining driving factors of streamflow based on big data analytics has been established.Aiming at dealing with the frequently changed statistic law and cause law of streamflow,a technical method of extracting driving factors of streamflow based on the combination of big data analytics and traditional data analytics methods was proposed.This method can remove the features that have no significant impact on streamflow,extract hidden features that affect streamflow,identify the driving relationship between features(such as hydrology,meteorology,and vegetation variables)and streamflow,and extract key driving factors of streamflow.Therefore,this method laid a data foundation for the construction of adaptive streamflow forecasting based on machine learning.3.A method for constructing streamflow forecasting models based on machine learning has been proposed.Due to the frequent changes of driving factors of streamflow,the structure and parameters of established models is inapplicable for future streamflow forecasting.Therefore,machine learning technology is introduced into to construct an establishment method for streamflow forecasting.This construction method realizes the self-learning of streamflow forecasting models and can effectively responds to environment and decision-demand change,laying a methodological foundation for establishing streamflow forecasting models.However,the black-box nature of machine learning models is a justifiable barrier of understanding these models and making decisions in water resource management.Hence,the interpretation of streamflow forecasting models was realized,which laid an application foundation for adaptive streamflow forecasting.4.Machine-learning-based streamflow forecasting models using statistic law of streamflow has been established.Aiming at adaptive streamflow forecasting in area lack meteorological and underlying surface observations,a procedure of constructing adaptive streamflow forecasting models based on statistic law of streamflow was proposed.Based on this procedure,streamflow forecasting models at different spatial and temporal scales were established using historical streamflow and based on signal processing algorithms and machine learning algorithms.Additionally,comparative evaluations of multiple models were carried out to prove this procedure avoids the use of future information,effectively reduces the influence of the boundary effect caused by signal processing algorithms.This procedure laid a model foundation for performing adaptive streamflow forecasting.5.Machine-learning-based streamflow forecasting models using cause law of streamflow has been established.Aiming at the impact of climate change and the evolution of the underlying surface environment on the formation process of streamflow,a procedure for constructing adaptive streamflow forecasting models based on the cause law of streamflow was proposed.Based on this procedure,streamflow forecasting models at different spatial and temporal scales were established using data including historical streamflow,meteorological as well as underlying surface features,and machine learning algorithms.Additionally,comparative evaluations of multiple models were carried out to prove this procedure not only effectively utilizes the cause-law information of streamflow,but also realizes the correction of prediction results based on historical streamflow information.This procedure laid a model foundation for the adaptive streamflow forecasting.6.An adaptive streamflow forecasting system has been designed and implemented.Aiming at the problem that adaptive streamflow forecasting needs process-based decision support,an adaptive streamflow forecasting system was designed using the pattern of "platform+content+service".Based on integrated support platform for knowledge visualization,a streamflow forecasting database,a model and method component library and a knowledge map library were constructed,which can help to realize the separate management of the streamflow forecasting data,models,and forecasting services.By integrating the knowledge map and components,the rapid construction of adaptive streamflow forecasting system and continuous adaptive streamflow forecasting services were realized.Through the continuous integration and application of the system,the simulation of streamflow forecasting and water dispatch for the Huangjinxia Reservoir of HANJIANG-TO-WEIHE RIVER VALLEY WATER DIVERSION PROJECT was performed,and the adaptive streamflow forecasting was finally realized.