Research On Coal Mine Water Inrush Multi-level Forecasting Method Based On Online Self-adaptive Evolutionary Extreme Learning Machine

Water inrush, which causes the most of the direct economic losses among all the mine disasters, is one of serious casualty accidents in China. So researches on both mine water inrush mechanism and an accurate and rapid water inrush forecasting model are needed to guarantee the safe production in coal mine.Water inrush refers to many factors such as mining conditions, hydrogeological conditions, rock properties, etc. Complex nonlinear relationship between the above factors and water inrush makes it difficult to construct an accurate and rapid water inrush forecasting model by traditional mathematical theories. Through the existing researches of coal mine water burst forecast methods, both the advantages and disadvantages of the methods are analyzed. Then, a new single layer feed-forward neural network named online self-adaptive evolutionary extreme learning machine(Online Sa E-ELM) is proposed to construct a multi-Level forecasting model for coal mine water inrush.Most of the exsiting researches are focused on the forecast of exsistence rather than the level of water inrush. However, the damage and response measures are different in different water inrush level. Therefore, water inrush is divided into four types in this thesis based on the maximum water inrush by national standard, such that pretreatment can be made for different water inrush situations through the multi-level forecasting model. Since the time changes, old forecast model may not be able to meet the needs of the new forecasting in practice. For Online Sa E-ELM, the output weights will be changed online according to the new inputs, then the network structure will be updated. The algorithm shows great advantage to solve multi-classification problems for the characteristics of good accuracy, generalization performance, rate and so on,By analyzing the mechanism and influencing factors of water inrush, main factors are chosen. Based on Online Sa E-ELM, a multi-level forecasting model is constructed on the platform of MATLAB according to the training and testing samples, which are provided from the historical water inrush data in Huaibei China. Then the network structure is determined according to the adjustment of the network parameters including the number of nodes in the hidden layer and excitation function through experiments. At last, a comparison between the results of the proposed forecast model and the results of the existing methods is presented to show the advantages on precision and speed of the forecast proposed in this thesis.