Intelligent prediction method of coal mine water inflow quantity by fusing multi-dimensional factors and deep time sequence model

Abstract

The present application introduces a kind of intelligent prediction method of coal mine water inflow quantity fusing multi-dimensional factors and deep time sequence model, it is related to wisdom mine and industrial big data analysis and prediction technical field, specific steps are: S100: the acquisition and feature engineering of multi-dimensional heterogeneous data;S200: data preprocessing and space-time alignment based on digital elevation model DEM;S300: feature dynamic weighted fusion based on gated attention mechanism;S400: water inflow quantity inference prediction based on deep time sequence model;S500: online closed-loop self-correction based on reinforcement learning.The present application compared with traditional single data source method, prediction accuracy is improved;Through digital elevation model, the space-time alignment relationship of surface water to downhole catchment point conduction is constructed, the prediction result is more in line with objective law;The unique reinforcement learning self-correction mechanism makes the model can learn and optimize independently according to real-time feedback, realizes the continuous evolution effect of more and more accurate, without manual intervention can adapt to the dynamic change of mine.

Description

Technical Field

[0001] This invention relates to the field of smart mining and industrial big data analysis and prediction technology, and in particular to an intelligent prediction method for coal mine water inflow that integrates multi-dimensional factors and deep time series models. Background Technology

[0002] Predicting coal mine water inflow is the core and prerequisite for water hazard prevention. Existing prediction methods have the following inherent defects, resulting in poor prediction performance in complex and variable mine environments: 1. Insufficient input dimensions, failing to reflect complex causes: Traditional methods mostly rely on only a few time series variables such as historical rainfall and water inflow for modeling, completely ignoring other factors that have a significant impact on water inflow; for example, engineering factors: the continuous progress of mining operations, geological factors: the distribution and water conductivity of geological faults, environmental factors: changes in surface vegetation and soil moisture content, etc. These multi-dimensional dynamic factors jointly determine the final form of water inflow. Existing models lack the ability to effectively integrate such data, resulting in systematic biases in their prediction results.

[0003] 2. Simple model structure, unable to capture nonlinear relationships: mostly adopts linear regression or traditional time series models, such as ARIMA. These models are difficult to capture such a complex, high-delay, and highly nonlinear dynamic process as mine water inrush.

[0004] 3. Lack of physical mechanism constraints, data-driven methods are prone to distortion: Existing methods are usually purely data-driven and do not take the basic laws of hydrogeology, such as the time delay effect of the transmission of surface rainfall to underground water inflow, as the prior knowledge or hard constraints of the model. This makes it easy for the model to produce absurd prediction results that violate physical laws when the training data is insufficient or the operating conditions change abruptly.

[0005] 4. Poor adaptability and inability to evolve online: Once the model is trained, its parameters are basically fixed, making it unable to learn and optimize online and automatically based on the latest predictive performance and changes in operating conditions. This offline training and online use model makes it unsuitable for the continuously dynamic system of a mine.

[0006] Therefore, there is an urgent need in this field for a novel prediction method that must: first, integrate heterogeneous data from multiple dimensions such as geology and engineering; second, employ advanced models that can deeply understand time series dependencies; third, respect and utilize the physical laws of hydrological conduction; and fourth, possess the ability to self-correct and continuously evolve online. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent prediction method for coal mine water inflow that integrates multidimensional factors and deep time series models. It adopts a combination of multidimensional feature engineering and physical time delay compensation to achieve a closed loop of advanced deep time series models and reinforcement learning adaptive correction, resulting in high prediction accuracy.

[0008] The technical solution adopted in this invention is: an intelligent prediction method for coal mine water inflow that integrates multi-dimensional factors and a deep time series model, characterized by the following specific steps: S100: Acquisition and Feature Engineering of Multidimensional Heterogeneous Data; The data receiving and feature engineering module acquires the raw information of the following four types of data from multiple heterogeneous data sources, and performs unified quantization, structuring and feature engineering processing on the data to obtain standardized input for modeling; Historical time series data includes historical water inflow and rainfall; standardized time series features are constructed by unifying time granularity and calculation units; Engineering factor data includes the daily advance of the mining face and the cumulative change in the volume of the goaf, quantifying engineering records and spatial geometric information into calculable numerical characteristics. Geological data include the density of geological faults along the working face, the distribution map of rock strata permeability coefficients, fault displacement, and the shortest distance from the current location of the working face to the fault. Structural geological features are formed through spatial statistics and parameter mapping methods. Environmental and weather factor data include rainfall and air pressure monitored by real-time rain sensors, as well as future rainfall and air pressure obtained through API, and are converted into a numerical feature sequence consistent with the forecast timeline.

[0009] S200: Data preprocessing and spatiotemporal alignment based on the Digital Elevation Model (DEM); Based on the four types of multidimensional features constructed in step S100, outlier removal, missing data completion, and normalization are performed on the feature data to eliminate the influence of dimensional differences and data noise on model training. At the same time, the mining area digital elevation model is introduced to physically model external environmental factors with spatial transmission and time delay effects. The hydrological transmission time delay parameter τ of external environmental factors from the surface to the underground water catchment point is analyzed and calculated. Based on the hydrological transmission time delay parameter τ, the time series of the corresponding external factors are shifted forward and aligned to obtain four types of complete multidimensional feature sequences; achieving causal consistency of multidimensional features in the temporal and spatial dimensions.

[0010] Specifically, in step S200, the step of using a digital elevation model to perform hydrological analysis and calculate the hydrological conduction time delay parameter τ is as follows: First, identify the main surface water catchment areas and extract the slope runoff path length based on the mining area's digital elevation model. With average slope In conjunction with rainfall intensity characteristics, the surface runoff conduction time from the catchment area to the infiltration area was calculated. The calculation formula is as follows: ;in, This indicates the peak flow rate formed by the confluence of runoff on the slope. The rainfall intensity index indicates the intensity of precipitation per unit time. The attenuation index indicates the attenuation effect of rainfall intensity in a region as rainfall duration increases; Second, by combining geological profile maps and borehole data of the mining area, densely fractured zones and water-conducting faults are identified as dominant conduction channels for rainwater infiltration, and the equivalent path of groundwater conduction is determined. Third, based on the identified dominant conduction pathways, combined with hydrogeological parameters: rock permeability coefficient The degree of fracture development or equivalent porosity Aquifer thickness parameters and hydraulic gradient Darcy's law was used to calculate the underground seepage conduction time of groundwater from the surface to the monitoring point in the well. The calculation formula is as follows: Where i represents the i-th dominant seepage channel, L i This indicates the equivalent seepage path length of the corresponding channel. The range of values ​​for the contribution weights of different seepage channels to overall hydrological conduction. ; Fourth, the conduction time of surface runoff and underground seepage conduction time The total hydrological transmission time delay τ is obtained through comprehensive calculation; Wherein, λ is the underground seepage influence coefficient, and its value range is... .

[0011] S300: Feature dynamic weighted fusion based on gated attention mechanism; The preprocessed and aligned multidimensional feature sequence is input into the fusion module based on gated attention mechanism; Based on the traditional attention calculation results, the fusion module introduces a gated mechanism for different attention heads, and achieves dynamic weighting of the contribution of different feature channels and different time steps by applying nonlinear gated adjustment to the attention output, so as to obtain the fused feature sequence.

[0012] Specifically, the feature dynamic weighted fusion method in step S300 is as follows: The gating attention-based fusion module receives four types of feature sequences after spatiotemporal alignment in step S200: historical time series data stream X1, engineering factor data stream X2, geological factor data stream X3, and environmental weather data stream X4. First, feature mapping: the four feature sequences X1, X2, X3, and X4 are mapped through learnable linear transformations to obtain the query vector Q. i Key vector K i Sum vector V i : , , , ;in, The learnable parameter matrix; Second, calculate the attention score; within each attention head, calculate the attention score based on the query vector and key vector, and normalize it using the softmax function to obtain the attention weights for each feature class: , ; Where h represents the attention head index, d is the feature dimension, and the softmax function ensures that the sum of all weights is 1; Third, single-head dynamic weighted fusion; based on the attention weights, the feature vectors of various types are weighted and summed to obtain the fused feature vector of the h-th attention head: ; Fourth, head-level sigmoid gating modulation; to further suppress redundant information and enhance the stability of the model during long-term sequence modeling, a head-level sigmoid gating mechanism is introduced, which modulates the output of each attention head. Construct independent gate functions: The output of this attention head is modulated by element-wise multiplication: ,in, Let h be the learnable gating parameters corresponding to the h-th attention head. It is the Sigmoid activation function. This represents element-wise multiplication; Fifth, multi-head attention fusion and linear mapping; the results of gating and modulating the outputs of all attention heads in the current time step are concatenated and then processed through a learnable linear mapping matrix. Dimension mapping is performed on the concatenated feature vectors to obtain the fused features at the current time step: , where H represents the total number of attention heads; Sixth, construct the fused feature sequence; repeat steps one through five of the gated multi-head attention fusion process for each time step in the time series to finally obtain the fused feature sequence. , which serves as the input for subsequent deep time series models, where T represents the length of the time series.

[0013] S400: Water inflow inference prediction based on deep time series model; The fused feature sequence obtained in step S300 is input into a pre-deployed deep time series model for inference calculation, and the predicted mine water inflow for a future period is output; The deep time series model adopts a sequence modeling network structure based on self-attention mechanism, including an encoding module and a decoding module; The encoding module consists of multiple stacked encoding layers, each layer including an extended multi-head attention mechanism and a feedforward network, used to extract time-dependent features from the input feature sequence and generate a high-dimensional representation; The decoding module consists of multiple stacked decoding layers, each layer containing self-attention, multi-head cross-attention and a feedforward network, used to gradually generate a sequence of predicted water inflow values ​​based on the high-dimensional representation output by the encoding module.

[0014] Specifically, in step S400, the output of each attention in the encoding and decoding layers is calculated using the following formula: The formula for calculating attention weights using the self-attention mechanism is as follows: ; in, This represents the query matrix, where each row represents the query vector for each feature value in the time series. This represents the key matrix, where each row represents the key vector for each feature value in the time series. This represents the transpose of the key matrix. The activation function transforms the attention score matrix into probabilities, ensuring that the sum of the scores equals 1. Representing the key vector, the dot product value can become very large as the vector dimension increases, leading to... Small gradients occur during the operation, so this value is used to balance the attention score; After obtaining the attention weights, the weighted sum of the value vectors is further calculated to obtain the output of each attention head, as shown in the following formula: ; in, The output of the attention head represents a new representation of each feature value in the time series, weighted by the relationship between consecutive time steps. This represents a value matrix, where each row represents the value vector of each feature in the time series; Multiple attention heads independently learn the different temporal dependencies of sequences, and the outputs are concatenated and linearly mapped to serve as the final representation of the encoding or decoding layer, as shown in the following formula: ;in, This represents the output of multiple self-attention heads. Represents the weight matrix. This represents a matrix concatenation operation; After the encoding and decoding modules complete their layer-by-layer calculations, the model outputs the final sequence of predicted water inflow values.

[0015] S500: Online closed-loop self-correction based on reinforcement learning; The predicted water inflow obtained in step S400 is compared with the actual water inflow collected by downhole sensors in real time, the deviation between the predicted result and the actual observed value is calculated, and an online self-correction mechanism is constructed to realize the dynamic optimization of the depth time series model.

[0016] Specifically, in step S500, the online self-correction process is modeled as a reinforcement learning problem, where the agent is the online correction control module, aiming to dynamically correct the deep temporal model running in real time; the action is to adjust the parameters of the feature fusion module in step S300 based on the prediction deviation, or to fine-tune the network weights of the deep temporal model in step S400; the reward function is defined as the negative value of the prediction deviation, that is, the smaller the deviation, the greater the reward; through continuous online interaction, the agent learns a set of optimal strategies, enabling it to automatically and predictively adjust the model based on real-time prediction performance, ultimately achieving online self-correction of the deep temporal model; specifically: State Construction: During the real-time operation of the system, the online correction control module constructs the current state s in each update cycle. t Its state includes at least the prediction deviation sequence at the current time and within the historical time window, the weight parameters in the current feature fusion module, the network parameters of the current deep time series model, and external environment state information; Action selection: The online correction control module is based on the current state s t Choose an action 'a' from the action space according to the Q-learning strategy. t The actions include adjusting the weight parameters of the feature fusion module and fine-tuning the network parameters of the deep temporal model; Action execution and model update: Execute action a t The parameters of the corresponding updated feature dynamic weighted fusion module or the parameters of the deep time series model are fine-tuned, and new inflow prediction results are generated based on the updated model. Environmental feedback and reward calculation: The updated prediction results are compared with the actual water inflow collected by downhole sensors, and the prediction deviation e is calculated. t And based on this, construct instant rewards; State transition: After the action is executed and the prediction is completed, the system enters a new motion state s. t+1 ; Q-value function update: The online correction control module updates the state-action value function based on the Q-learning algorithm, and the update rules are as follows: ;in, Indicates the current state Take action below of value, The learning rate determines the speed at which new information is updated. Indicates the state Take action below The instant reward obtained afterward This serves as a discount factor, measuring the impact of future rewards. This indicates the next state. The largest of the following choices value; The above steps are continuously executed in a loop during system operation; through continuous online interaction and Q-value updates, the agent gradually learns a set of optimal strategies, ultimately achieving online self-correction of the surge depth time series model.

[0017] Due to the adoption of the technical solution described above, the present invention has the following advantages: 1. Multi-dimensional information fusion significantly improves prediction accuracy: By integrating multi-dimensional heterogeneous data from engineering, geology, and environment, the model can fully understand the complex causal mechanism of water inrush formation. Compared with traditional single data source methods, the prediction accuracy is improved by 20-30%.

[0018] 2. Deep integration of physical mechanism constraints and data-driven approach: By constructing the spatiotemporal alignment relationship of surface water conduction to underground water collection points through digital elevation model, the physical time delay law of hydrological conduction is introduced into the model, avoiding the blindness and uninterpretability of pure data-driven methods, and the prediction results are more in line with objective laws.

[0019] 3. The present invention adopts a deep temporal modeling network structure, which can effectively characterize the long-term dependence and nonlinear evolution characteristics of water inflow over time. It has stronger adaptability and robustness to water inflow processes with obvious hysteresis effects and sudden changes in mining conditions.

[0020] 4. Online continuous evolution based on reinforcement learning: The unique reinforcement learning self-correction mechanism enables the model to learn and optimize autonomously based on real-time feedback, achieving a continuous evolution effect that becomes more accurate with use, and can adapt to the dynamic changes in the mine without human intervention. Attached Figure Description

[0021] Figure 1 This is the overall flowchart of the present invention.

[0022] Figure 2 This is a schematic diagram of the functional modules of the present invention.

[0023] Figure 3This is a schematic diagram illustrating the spatiotemporal alignment principle based on the digital elevation model in an embodiment of the present invention.

[0024] In the diagram: 101-Real-time rain sensor, 102-Historical database, 103-Third-party weather API, 104-Mine area hydrological sensor, 105-Engineering data acquisition system, 106-Geological exploration database; 200-Edge computing unit, 201-Data receiving and feature engineering module, 202-Spatiotemporal alignment module, 203-Gated attention-based fusion module, 204-Deep temporal model, 205-Reinforcement learning self-correction module; 301 - Schematic diagram of hydrological conduction; 302 - Schematic diagram of surface rainfall variation over time; 303 - Schematic diagram of underground water inflow variation over time; 310-Surface, 311-Rainfall events; 312-Surface runoff and infiltration, 320-Rock strata and fissures, 330-Ground goaf or aquifer, 340-Underground tunnels and monitoring points; 351 - Surface rainfall time curve, 352 - Downhole water inflow time curve, t - time, Q - flow rate, τ - hydrological conduction time delay parameter; Detailed Implementation

[0025] The present invention will be further explained and described below with reference to the accompanying drawings and embodiments. However, this should not be construed as limiting the scope of protection of the present invention. The purpose of disclosing the present invention is to protect all technical improvements within the scope of the present invention.

[0026] The intelligent prediction method for coal mine water inflow that integrates multidimensional factors and deep time series models of the present invention is executed on an edge computing unit 200 deployed in a mine surface monitoring center. The edge computing unit 200 is typically an industrial server equipped with a GPU accelerator card to support efficient inference of deep learning models.

[0027] Combined with appendix Figure 1-3 The method for intelligent prediction of coal mine water inflow, which integrates multi-dimensional factors and a deep time series model, is shown below. The specific steps are as follows: S100: Acquisition and feature engineering of multidimensional heterogeneous data; The data receiving and feature engineering module 201 acquires the raw information of the following four types of data from multiple heterogeneous data sources, and performs unified quantization, structuring and feature engineering processing on the data to provide standardized input for subsequent modeling; Historical time series data: retrieve historical water inflow and rainfall data from the past 6-12 months from historical database 102; construct standardized time series features by unifying time granularity and calculation units.

[0028] Engineering factor data: Real-time daily advance of the mining face and cumulative changes in goaf volume are acquired from the engineering data acquisition system 105. By analyzing the spatial structure information in the mine's digital engineering drawings, engineering records and spatial geometric information are quantified into calculable numerical features; the newly added goaf volume (unit: cubic meters / day) is automatically calculated. This feature reflects the changes in underground space caused by mining activities, directly affecting the confluence path and inflow of groundwater.

[0029] Geological factor data: Geological structural information along the working face is read from the geological exploration database 106; including geological factor data such as the density of geological faults passed by the working face, the distribution map of rock strata permeability coefficient, fault displacement, and the shortest distance from the current position of the working face to the fault. Structural geological features are formed through spatial statistics and parameter mapping methods. The formula for calculating the fault influence index at the working face is as follows: ;in, The fault impact index of the working face. For the first The elevation difference between the fault line and the current working face. For the current working face to the first The shortest distance of the fault, The constant is usually 1 or a very small value; the working face fault influence index comprehensively considers the number, scale, dip angle and water conductivity of faults, and quantifies the degree of influence of geological structure on water inflow; the spatial distribution data of rock strata permeability coefficient is used as an auxiliary feature.

[0030] Environmental and weather data include minute-level rainfall and air pressure monitored by real-time rain sensors 101, as well as 72-hour rainfall forecasts, air pressure, and humidity meteorological elements obtained from third-party weather APIs 103, which are then converted into numerical feature sequences consistent with the forecast timeline. This data is crucial for predicting short-term, sudden water inrushes.

[0031] S200: Data preprocessing and spatiotemporal alignment based on the Digital Elevation Model (DEM); Based on the four types of multidimensional features constructed in step S100, outlier removal, missing data completion, and normalization are performed on the feature data to eliminate the influence of dimensional differences and data noise on model training. At the same time, the mining area digital elevation model is introduced to physically model external environmental factors with spatial transmission and time delay effects. The hydrological transmission time delay parameter τ of external environmental factors from the surface to the underground water catchment point is analyzed and calculated. Based on the hydrological transmission time delay parameter τ, the time series of the corresponding external factors are shifted forward and aligned to obtain four types of complete multidimensional feature sequences; achieving causal consistency of multidimensional features in the temporal and spatial dimensions.

[0032] like Figure 3 ,use Figure 2The spatiotemporal alignment module 202 performs this step; this module is pre-loaded with 1:5000 scale digital elevation model data of the mining area; such as Figure 3 As shown, when rainfall event 311 occurs at the surface 310 at time t0, the rainwater does not reach the underground roadway 340 instantly; it needs to go through a complete hydrological conduction process, including surface runoff and infiltration 312, infiltration downward along rock fissures 320, collection in the goaf or aquifer 330, and finally entering the underground roadway and monitoring point 340; this process has a significant time lag τ.

[0033] The steps for calculating the hydrological conduction time delay parameter τ using a digital elevation model for hydrological analysis are as follows: First, identify the main surface water catchment areas and extract the slope runoff path length based on the mining area's digital elevation model. With average slope In conjunction with rainfall intensity characteristics, the surface runoff conduction time from the catchment area to the infiltration area was calculated. The calculation formula is as follows: ;in, This indicates the peak flow rate formed by the confluence of runoff on the slope. The rainfall intensity index indicates the intensity of precipitation per unit time. The attenuation index indicates the attenuation effect of rainfall intensity in a region as rainfall duration increases; Second, by combining geological profile maps and borehole data of the mining area, densely fractured zones and water-conducting faults are identified as dominant conduction channels for rainwater infiltration, and the equivalent path of groundwater conduction is determined. Third, based on the identified dominant conduction pathways, combined with hydrogeological parameters: rock permeability coefficient The degree of fracture development or equivalent porosity Aquifer thickness parameters and hydraulic gradient Darcy's law was used to calculate the underground seepage conduction time of groundwater from the surface to the monitoring point in the well. The calculation formula is as follows: Where i represents the i-th dominant seepage channel, L i This indicates the equivalent seepage path length of the corresponding channel. This represents the weight of the contribution of different seepage channels to the overall hydrological conduction, with a value range of... ; Fourth, the conduction time of surface runoff and underground seepage conduction time The total hydrological transmission time delay τ is obtained through comprehensive calculation; Wherein, λ is the underground seepage influence coefficient, and its value range is... .

[0034] from Figure 3 The comparison of the time curves below shows that the peak of the surface rainfall time curve 351 occurs at time t0, while the peak of the underground water inflow time curve 352 occurs at time t0+τ.

[0035] In order for the model to correctly learn the causal relationship between rainfall and water inrush, this step forwards the timestamp t of all external data related to surface rainfall (including data from real-time rain sensor 101 and forecast data from third-party weather API 103): t' = t + τ; after alignment, the rainfall data at time t and the water inrush data at time t+τ achieve a physical correspondence in the time dimension.

[0036] S300: Feature dynamic weighted fusion based on gated attention mechanism; The preprocessed and aligned multidimensional feature sequence is input into the fusion module based on gated attention mechanism; Based on the traditional attention calculation results, the fusion module introduces a gated mechanism for different attention heads, and achieves dynamic weighting of the contribution of different feature channels and different time steps by applying nonlinear gated adjustment to the attention output, so as to obtain the fused feature sequence. The feature dynamic weighted fusion method in step S300 is as follows: The gating attention-based fusion module 203 receives four types of feature sequences after spatiotemporal alignment in step S200: historical time series data stream 401X1, engineering factor data stream 402X2, geological factor data stream 403X3, and environmental weather 404X4. First, feature mapping: the four feature sequences X1, X2, X3, and X4 are mapped through learnable linear transformations to obtain the query vector Q. i Key vector K i Sum vector V i : , , , ;in, The learnable parameter matrix; Second, calculate the attention score; within each attention head, calculate the attention score based on the query vector and key vector, and normalize it using the softmax function to obtain the attention weights for each feature class: , ; Where h represents the attention head index, d is the feature dimension, and the softmax function ensures that the sum of all weights is 1; that is... ; Third, single-head dynamic weighted fusion; based on the attention weights, the feature vectors of various types are weighted and summed to obtain the fused feature vector of the h-th attention head: ; Fourth, head-level sigmoid gating modulation; to further suppress redundant information and enhance the stability of the model during long-term sequence modeling, a head-level sigmoid gating mechanism is introduced, which modulates the output of each attention head. Construct independent gate functions: The output of this attention head is modulated by element-wise multiplication: ,in, Let h be the learnable gating parameters corresponding to the h-th attention head. It is the Sigmoid activation function. This represents element-wise multiplication; Fifth, multi-head attention fusion and linear mapping; the results of gating and modulating the outputs of all attention heads in the current time step are concatenated and then processed through a learnable linear mapping matrix. Dimension mapping is performed on the concatenated feature vectors to obtain the fused features at the current time step: , where H represents the total number of attention heads; Sixth, construct the fused feature sequence; repeat steps one through five of the gated multi-head attention fusion process for each time step in the time series to finally obtain the fused feature sequence. , which serves as the input for subsequent deep time series models, where T represents the length of the time series.

[0037] S400: Water inflow inference prediction based on deep time series model; The fused feature sequence obtained in step S300 is input into the pre-deployed deep time series model 204 for inference calculation, and the predicted mine water inflow results for a future period are output; The deep time series model adopts a sequence modeling network structure based on self-attention mechanism, including an encoding module and a decoding module; The encoding module consists of multiple stacked encoding layers, each layer including an extended multi-head attention mechanism and a feedforward network, used to extract time-dependent features from the input feature sequence and generate a high-dimensional representation; The decoding module consists of multiple stacked decoding layers, each layer including self-attention, multi-head cross-attention and a feedforward network, used to gradually generate the water inflow prediction value sequence at the corresponding time based on the high-dimensional representation output by the encoding module.

[0038] In the encoding and decoding layers, the output of each attention level is calculated using the following formula: The formula for calculating attention weights using the self-attention mechanism is as follows: ; in, This represents the query matrix, where each row represents the query vector for each feature value in the time series. This represents the key matrix, where each row represents the key vector for each feature value in the time series. This represents the transpose of the key matrix. The activation function transforms the attention score matrix into probabilities, ensuring that the sum of the scores equals 1. Representing the key vector, the dot product value can become very large as the vector dimension increases, leading to... Small gradients occur during the operation, so this value is used to balance the attention score; After obtaining the attention weights, the weighted sum of the value vectors is further calculated to obtain the output of each attention head, as shown in the following formula: ; in, The output of the attention head represents a new representation of each feature value in the time series, weighted by the relationship between consecutive time steps. This represents a value matrix, where each row represents the value vector of each feature in the time series; Multiple attention heads independently learn the different temporal dependencies of sequences, and the outputs are concatenated and linearly mapped to serve as the final representation of the encoding or decoding layer, as shown in the following formula: ;in, This represents the output of multiple self-attention heads. Represents the weight matrix. This represents a matrix concatenation operation; After the encoding and decoding modules complete their layer-by-layer calculations, the model outputs the final sequence of predicted water inflow values.

[0039] S500: Online closed-loop self-correction based on reinforcement learning; The predicted water inflow obtained in step S400 is compared with the actual water inflow collected by downhole sensors in real time, the deviation between the predicted result and the actual observed value is calculated, and an online self-correction mechanism is constructed to realize the dynamic optimization of the depth time series model.

[0040] Specifically, the online self-correction process is modeled as a reinforcement learning problem, where the agent acts as an online correction control module, aiming to dynamically correct the deep temporal model in real-time operation. Actions include adjusting the parameters of the feature fusion module in step S300 based on the prediction bias, or fine-tuning the network weights of the deep temporal model in step S400. The reward function is defined as the negative value of the prediction bias; that is, the smaller the bias, the greater the reward. Through continuous online interaction, the agent learns a set of optimal strategies, enabling it to automatically and proactively adjust the model based on real-time prediction performance, ultimately achieving online self-correction of the deep temporal model. Specifically: State Construction: During the real-time operation of the system, the online correction control module constructs the current state s in each update cycle. t Its state includes at least the prediction deviation sequence at the current time and within the historical time window, the weight parameters in the current feature fusion module, the network parameters of the current deep time series model, and external environment state information; Action selection: The online correction control module is based on the current state s t Choose an action 'a' from the action space according to the Q-learning strategy. t The actions include adjusting the weight parameters of the feature fusion module and fine-tuning the network parameters of the deep temporal model; Action execution and model update: Execute action a t The parameters of the corresponding updated feature dynamic weighted fusion module or the parameters of the deep time series model are fine-tuned, and new inflow prediction results are generated based on the updated model. Environmental feedback and reward calculation: The updated prediction results are compared with the actual water inflow collected by downhole sensors, and the prediction deviation e is calculated. t And based on this, construct instant rewards; State transition: After the action is executed and the prediction is completed, the system enters a new motion state s. t+1 ; Q-value function update: The online correction control module updates the state-action value function based on the Q-learning algorithm, and the update rules are as follows: ;in, Indicates the current state Take action below of value, The learning rate determines the speed at which new information is updated. Indicates the state Take action below The instant reward obtained afterward This serves as a discount factor, measuring the impact of future rewards. This indicates the next state. The largest of the following choices value; The above steps are continuously executed in a loop during system operation; through continuous online interaction and Q-value updates, the agent gradually learns a set of optimal policies, ultimately achieving online self-correction of the deep time series model.

[0041] The parts of this invention not described in detail are prior art.

[0042] The embodiments selected herein for the purpose of disclosing the inventive objectives are currently considered suitable; however, it should be understood that the invention is intended to include all variations and modifications of the embodiments that fall within the scope of this concept and invention.

Claims

1. A method for intelligent prediction of coal mine water inflow integrating multi-dimensional factors and deep time series models, characterized in that: The specific steps are as follows: S100: Acquisition and feature engineering of multidimensional heterogeneous data; The data receiving and feature engineering module acquires the raw information of the following four types of data from multiple heterogeneous data sources, and performs unified quantization, structuring and feature engineering processing on the data; Historical time series data includes historical water inflow and rainfall; standardized time series features are constructed by unifying time granularity and calculation units; Engineering factor data includes the daily advance of the mining face and the cumulative change in the volume of the goaf, quantifying engineering records and spatial geometric information into calculable numerical characteristics. Geological data include the density of geological faults along the working face, the distribution map of rock strata permeability coefficients, fault displacement, and the shortest distance from the current location of the working face to the fault. Structural geological features are formed through spatial statistics and parameter mapping methods. Environmental and weather factor data include rainfall and air pressure monitored by real-time rain sensors, as well as future rainfall and air pressure obtained through API, and are converted into a numerical feature sequence consistent with the forecast timeline; S200: Data preprocessing and spatiotemporal alignment based on the digital elevation model (DEM); Based on the four types of multidimensional features constructed in step S100, outlier removal, missing data completion, and normalization are performed on the feature data. At the same time, the mining area digital elevation model is introduced to physically model the external environmental factors with spatial transmission and time delay effects. The hydrological transmission time delay parameter τ of the external environmental factors from the surface to the underground water collection point is analyzed and calculated. Based on the hydrological transmission time delay parameter τ, the time series of the corresponding external factors are shifted forward and aligned to obtain four types of complete multidimensional feature sequences. S300: Dynamic weighted fusion of features based on gated attention mechanism; inputting preprocessed and aligned multidimensional feature sequences into the fusion module based on gated attention mechanism; Based on the traditional attention calculation results, this fusion module introduces a gating mechanism for different attention heads. By applying nonlinear gating adjustment to the attention output, it achieves dynamic weighting of the contributions of different feature channels and different time steps, thus obtaining a fused feature sequence. S400: Inflow inference prediction based on deep time series model; The fused feature sequence obtained in step S300 is input into a pre-deployed deep time series model for inference calculation, outputting the predicted mine water inflow for a future period. The deep time series model adopts a sequence modeling network structure based on a self-attention mechanism, including an encoding module and a decoding module. The encoding module consists of multiple stacked encoding layers, each layer including an extended multi-head attention mechanism and a feedforward network, used to extract time-dependent features from the input feature sequence and generate a high-dimensional representation. The decoding module consists of multiple stacked decoding layers, each layer containing self-attention, multi-head cross-attention, and a feedforward network, used to gradually generate a sequence of predicted water inflow values ​​based on the high-dimensional representation output by the encoding module. S500: Online closed-loop self-correction based on reinforcement learning; The predicted water inflow obtained in step S400 is compared with the actual water inflow collected by downhole sensors in real time, the deviation between the predicted result and the actual observed value is calculated, and an online self-correction mechanism is constructed to realize the dynamic optimization of the depth time series model.

2. The intelligent prediction method for coal mine water inflow based on the integration of multidimensional factors and deep time series models according to claim 1, characterized in that: In step S200, the step of using a digital elevation model to perform hydrological analysis and calculate the hydrological conduction time delay parameter τ is as follows: First, identify the main surface water catchment areas and extract the slope runoff path length based on the mining area's digital elevation model. With average slope In conjunction with rainfall intensity characteristics, the surface runoff conduction time from the catchment area to the infiltration area was calculated. The calculation formula is as follows: ;in, This indicates the peak flow rate formed by the confluence of runoff on the slope. The rainfall intensity index indicates the intensity of precipitation per unit time. The attenuation index indicates the attenuation effect of rainfall intensity in a region as rainfall duration increases; Second, by combining geological profile maps and borehole data of the mining area, densely fractured zones and water-conducting faults are identified as dominant conduction channels for rainwater infiltration, and the equivalent path of groundwater conduction is determined. Third, based on the identified dominant conduction pathways, combined with hydrogeological parameters: rock permeability coefficient The degree of fracture development or equivalent porosity Aquifer thickness parameters and hydraulic gradient Darcy's law was used to calculate the underground seepage conduction time of groundwater from the surface to the monitoring point in the well. The calculation formula is as follows: Where i represents the i-th dominant seepage channel, L i This indicates the equivalent seepage path length of the corresponding channel. The range of values ​​for the contribution weights of different seepage channels to overall hydrological conduction. ; Fourth, the conduction time of surface runoff and underground seepage conduction time The total hydrological transmission time delay τ is obtained through comprehensive calculation; Wherein, λ is the underground seepage influence coefficient, and its value range is... .

3. The intelligent prediction method for coal mine water inflow based on the integration of multidimensional factors and deep time series models as described in claim 1, characterized in that: The feature dynamic weighted fusion method in step S300 is as follows: The gating attention-based fusion module receives four types of feature sequences after spatiotemporal alignment in step S200: historical time series data stream X1, engineering factor data stream X2, geological factor data stream X3, and environmental weather data stream X4. First, feature mapping: the four feature sequences X1, X2, X3, and X4 are mapped through learnable linear transformations to obtain the query vector Q. i Key vector K i Sum vector V i : , , , ;in, The learnable parameter matrix; Second, calculate the attention score; within each attention head, calculate the attention score based on the query vector and key vector, and normalize it using the softmax function to obtain the attention weights for each feature class: ， ； Where h represents the attention head index, d is the feature dimension, and the softmax function ensures that the sum of all weights is 1; Third, single-head dynamic weighted fusion; based on the attention weights, the feature vectors of various types are weighted and summed to obtain the fused feature vector of the h-th attention head: ; Fourth, head-level sigmoid-gated modulation; a head-level sigmoid gating mechanism is introduced to control the output of each attention head. Construct independent gate functions: The output of this attention head is modulated by element-wise multiplication: ,in, Let h be the learnable gating parameters corresponding to the h-th attention head. It is the Sigmoid activation function. This represents element-wise multiplication; Fifth, multi-head attention fusion and linear mapping; the results of gating and modulating the outputs of all attention heads in the current time step are concatenated and then processed through a learnable linear mapping matrix. Dimension mapping is performed on the concatenated feature vectors to obtain the fused features at the current time step: , where H represents the total number of attention heads; Sixth, construct the fused feature sequence; repeat steps one through five of the gated multi-head attention fusion process for each time step in the time series to finally obtain the fused feature sequence. , which serves as the input for subsequent deep time series models, where T represents the length of the time series.

4. The intelligent prediction method for coal mine water inflow based on the integration of multidimensional factors and deep time series model as described in claim 1, characterized in that: In step S400, the output of each attention point in the encoding and decoding layers is calculated using the following formula: The formula for calculating attention weights using the self-attention mechanism is as follows: ； in, This represents the query matrix, where each row represents the query vector for each feature value in the time series. This represents the key matrix, where each row represents the key vector for each feature value in the time series. This represents the transpose of the key matrix. For activation function, Represents the key vector; After obtaining the attention weights, the weighted sum of the value vectors is further calculated to obtain the output of each attention head, as shown in the following formula: ； in, The output of the attention head represents a new representation of each feature value in the time series, weighted by the relationship between consecutive time steps. This represents a value matrix, where each row represents the value vector of each feature in the time series; Multiple attention heads independently learn the different temporal dependencies of sequences, and the outputs are concatenated and linearly mapped to serve as the final representation of the encoding or decoding layer, as shown in the following formula: ;in, This represents the output of multiple self-attention heads. Represents the weight matrix. This represents a matrix concatenation operation; After the encoding and decoding modules complete their layer-by-layer calculations, the model outputs the final sequence of predicted water inflow values.

5. The intelligent prediction method for coal mine water inflow based on the integration of multidimensional factors and a deep time series model as described in claim 1, characterized in that: In step S500, the agent is an online correction control module; it performs online correction of the operating parameters of the deep temporal model, by adjusting the parameters of the feature fusion module in step S300 based on the prediction deviation, or by fine-tuning the network weights of the deep temporal model in step S400; the reward function is defined as the negative value of the prediction deviation, i.e., the smaller the deviation, the larger the reward; specifically: State Construction: During the real-time operation of the system, the online correction control module constructs the current state s in each update cycle. t Its state includes at least the prediction deviation sequence at the current time and within the historical time window, the weight parameters in the current feature fusion module, the network parameters of the current deep time series model, and external environment state information; Action selection: The online correction control module is based on the current state s t Choose an action 'a' from the action space according to the Q-learning strategy. t The actions include adjusting the weight parameters of the feature fusion module and fine-tuning the network parameters of the deep temporal model; Action execution and model update: Execute action a t The parameters of the corresponding updated feature dynamic weighted fusion module or the parameters of the deep time series model are fine-tuned, and new inflow prediction results are generated based on the updated model. Environmental feedback and reward calculation: The updated prediction results are compared with the actual water inflow collected by downhole sensors, and the prediction deviation e is calculated. t And based on this, construct instant rewards; State transition: After the action is executed and the prediction is completed, the system enters a new motion state s. t+1 ; Q-value function update: The online correction control module updates the state-action value function based on the Q-learning algorithm, and the update rules are as follows: ;in, Indicates the current state Take action below of value, Indicates the learning rate. Indicates the state Take action below The instant reward obtained afterward As a discount factor, This indicates the next state. The largest of the following choices value; The above steps are continuously executed in a loop during system operation; through continuous online interaction and Q-value updates, the agent gradually learns a set of optimal policies, ultimately achieving online self-correction of the deep time series model.

Images