Chained transmission prediction method and system for risk of small hydropower station group in watershed

Abstract

The invention relates to the technical field of hydropower station safety management, in particular to a watershed small hydropower station group risk chain transmission prediction method and system, and the method comprises the steps: obtaining real-time hydrometeorological data and historical operation data, constructing a layered embedded risk transmission prediction architecture, and constructing a virtual small hydropower station group complex network. And a multi-domain coupling model is established by using an embedded time sequence deep learning framework, the model processes multi-source heterogeneous data through a topology perception input mechanism, captures time sequence dynamic characteristics through a risk state memory mechanism, and generates a risk transmission prediction result through a multi-domain coupling output mechanism, so that the risk prediction accuracy is remarkably improved, and the risk prediction efficiency is improved. Compared with a traditional method, the method has the advantages that the prediction efficiency is improved by about 35%, particularly, the prediction effect in a complex chain transmission scene is more remarkable, meanwhile, the method can find potential risks 4-6 hours in advance, and sufficient preparation time is reserved for countermeasures.

Description

Technical Field

[0001] The present invention relates to the technical field of hydropower station safety management, and particularly to a method and system for predicting risk chain transmission of small hydropower station groups in a basin. Background Art

[0002] As an important part of renewable clean energy, small hydropower stations play an important role in China's energy structure. However, due to the large number, wide distribution, and decentralized management of small and medium-sized hydropower stations in the basin, they face various risks during operation, such as flood disasters, equipment failures, and ecological environment impacts, and these risks have obvious chain transmission characteristics within the basin.

[0003] In the prior art, the risk assessment of small hydropower stations usually adopts a single-station isolated analysis method, which is difficult to reflect the risk transmission relationship between stations. Some studies attempt to analyze the relationship between stations from the perspective of hydrological links, but most are limited to the analysis of a single risk type, such as only focusing on flood risk or only focusing on equipment failure risk, and it is difficult to depict the coupling transmission mechanism between different types of risks. In addition, the existing prediction methods generally have problems such as insufficient timeliness and low prediction accuracy, and cannot meet the actual needs of the safety management of small hydropower station groups in the basin.

[0004] With the development of artificial intelligence technology, deep learning has made remarkable progress in the field of time series prediction. However, how to organically combine complex network theory with deep learning technology to achieve accurate prediction of risk chain transmission of small hydropower station groups in the basin is still a technical problem to be solved urgently. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for predicting risk chain transmission of small hydropower station groups in a basin, aiming to solve the problem in the prior art that it is difficult to depict the chain transmission law of multiple types of risks of small hydropower station groups, and to achieve accurate prediction of the risk chain transmission of flood - equipment failure - ecological pollution.

[0006] The present invention proposes a method for predicting risk chain transmission of small hydropower station groups in a basin, including: Obtaining real-time hydrometeorological data and historical operation data of small hydropower station groups in the basin; including: Based on the real-time hydrometeorological data and historical operation data, constructing a hierarchical embedded risk transmission prediction architecture including a physical connection layer, a risk mapping layer, and a multi-domain coupling layer; According to the hierarchical embedded risk transmission prediction architecture, constructing a complex network of virtual small hydropower station groups; Based on the complex network of virtual small hydropower station groups, establishing a multi-domain coupling model of flood - equipment failure - ecological pollution using an embedded time series deep learning framework; including: Processing multi-source heterogeneous data using a topology-aware input mechanism; Capture the temporal dynamic characteristics through the risk state memory mechanism; Generate the risk transfer prediction result through the multi-domain coupling output mechanism; Output the risk transfer prediction result to achieve the prediction of the risk chain transfer of small hydropower station groups in the basin.

[0007] Preferably, the acquisition of the real-time hydrometeorological data and historical operation data of the small hydropower station group in the basin specifically includes: Collect the water level, flow rate, rainfall, and temperature data of the upstream and downstream basin power stations, with a sampling frequency of 15 minutes / time; Obtain the operation status data of key equipment such as generator sets, transformers, and gates of the small hydropower station group, with a sampling frequency of 1 minute / time; Collect water quality parameters such as dissolved oxygen, pH value, and turbidity of the water quality monitoring section, with a sampling frequency of 1 hour / time; Integrate the data of historical flood events, equipment failure records, and ecological pollution events for at least 5 years.

[0008] Preferably, the construction of the complex network of the virtual small hydropower station group specifically includes: Take each small hydropower station in the basin as a node and construct a node attribute vector, including geographical coordinates, installed capacity, regulation capacity, and equipment parameters; Establish a directed weighted edge based on the hydraulic connection of the river channel, and comprehensively consider the hydraulic transfer time, flow correlation degree, and geographical proximity for the edge weight; Calibrate the hydraulic connection parameters and edge weights between nodes according to the measured hydrological data; Use the community detection algorithm to identify subgroups of small hydropower stations with close connections and form a hierarchical network structure.

[0009] Preferably, the construction of the hierarchical embedded risk transfer prediction architecture including a physical connection layer, a risk mapping layer, and a multi-domain coupling layer specifically includes: In the physical connection layer, establish a node connection relationship based on hydraulic characteristics to capture the water flow dynamics transfer characteristics; In the risk mapping layer, construct a risk measurement space, define a risk state vector and a risk distance function to characterize the risk spatio-temporal heterogeneity; In the multi-domain coupling layer, establish a two-way coupling mechanism for flood-equipment failure, equipment failure-ecological pollution, and ecological pollution-flood risks to depict the interaction relationship between multiple types of risks.

[0010] Preferably, the use of the topology-aware input mechanism to process multi-source heterogeneous data specifically includes: Realize multi-source data fusion and unify the data formats and time tags of different sources and different sampling frequencies; Extract local topological features, medium-distance topological features, and global topological features to form a multi-scale topological feature representation; Identify short-term fluctuation features, medium-term trend features, and long-term cycle features to construct multi-scale time series features; Selectively focus on the feature changes of key nodes and key time periods through the attention mechanism.

[0011] Preferably, the capturing of time series dynamic characteristics through the risk state memory mechanism specifically includes: Construct a multi-scale state representation of single-station state representation, local network state representation, and global network state representation; Achieve the dynamic balance between short-term memory update and long-term memory maintenance, and retain key historical information; Establish a non-linear state transition mechanism based on threshold triggering to capture the sharp state changes caused by emergencies; Implement smooth transition processing before and after the state transition to ensure the continuity of the prediction results.

[0012] Preferably, the generating of the risk transfer prediction result through the multi-domain coupling output mechanism specifically includes: Decode the internal state into flood risk level, equipment failure probability, and ecological pollution index respectively; Generate short-term risk predictions for the next 24 hours and medium- and long-term risk trends for the next 7 days; Predict the time sequence, transfer intensity, and coupling effect of risk transfer between different stations; Determine the warning level, warning range, and suggestions for countermeasures based on the prediction results.

[0013] Preferably, it further includes the verification and optimization process of the risk chain transfer model: Verify the fitting ability of the model through the reproduction of historical events; Use the comparison between actual observations and predictions to analyze the sources of prediction errors and biases; Optimize the model parameters and structure based on the verification results; Establish a prediction accuracy evaluation mechanism, and the prediction accuracy reaches more than 90%.

[0014] Preferably, the application scenarios of the risk chain transfer prediction method for small hydropower station groups in the basin include: Flood risk warning during the flood season, detecting potential risks 4 - 6 hours in advance; Equipment failure prediction and preventive maintenance, reducing the failure rate by more than 30%; Ecological environmental pollution warning and formulation of protection measures, reducing the incidence of ecological pollution events by more than 50%; Collaborative scheduling optimization of small hydropower station groups in the basin, improving the water resource utilization efficiency by more than 15%.

[0015] A risk chain transmission prediction system for small hydropower station groups in a river basin for implementing the above method, comprising: A data acquisition module for acquiring real-time hydrometeorological data and historical operation data of small hydropower station groups in a river basin; A network construction module for constructing a complex network of virtual small hydropower station groups; A risk mapping module for realizing the characterization of risk spatio-temporal heterogeneity; A multi-domain coupling module for establishing a coupling relationship of flood - equipment failure - ecological pollution; A prediction and analysis module for generating risk transmission prediction results by using an embedded time series deep learning framework; A decision support module for generating early warning information and intervention measure suggestions based on the prediction results; A model optimization module for realizing model verification and parameter optimization.

[0016] By constructing a hierarchical embedded risk transmission prediction architecture, the present invention organically combines complex network theory with deep learning technology, and realizes the coupling analysis and transmission prediction of multiple types of risks, having the following beneficial effects: 1. Significantly improves the accuracy of risk prediction, about 35% higher than traditional methods, especially the prediction effect is more significant in complex chain transmission scenarios; 2. Extends the early warning time, can discover potential risks 4 - 6 hours in advance, leaving sufficient preparation time for response measures; 3. Improves the water resource utilization efficiency, through accurate prediction to guide the coordinated scheduling of power station groups, making the water resource utilization efficiency increase by more than 15%; 4. Enhances the ecological environment protection effect, through accurate prediction of pollution risks, reducing the occurrence rate of ecological pollution events by more than 50%; 5. Improves the scalability of the system, and the architecture design supports small hydropower station groups of different scales and types, with strong adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a schematic flow chart of the risk chain transmission prediction method for small hydropower station groups in the river basin of the present invention; Figure 2 is a schematic diagram of the hierarchical embedded risk transmission prediction architecture of the present invention; Figure 3 is a flow chart for constructing the complex network of virtual small hydropower station groups of the present invention; Figure 4 is a structural diagram of the embedded time series deep learning prediction framework of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] Please refer to the attached Figures 1 - 4, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0019] Referring to Figure 1 As shown, the risk chain transfer prediction method for small hydropower station groups in a basin proposed by the present invention includes the following steps: 1) Obtain the real-time hydrometeorological data and historical operation data of the small hydropower station group in the basin; 2) Based on the real-time hydrometeorological data and historical operation data, construct a hierarchical embedded risk transfer prediction architecture including a physical connection layer, a risk mapping layer, and a multi-domain coupling layer; 3) According to the hierarchical embedded risk transfer prediction architecture, construct a complex network of virtual small hydropower station groups; 4) Based on the complex network of virtual small hydropower station groups, establish a multi-domain coupling model of flood - equipment failure - ecological pollution using an embedded time-series deep learning framework; 5) Output the risk transfer prediction result to achieve the prediction of the risk chain transfer of the small hydropower station group in the basin.

[0020] Preferably, the obtaining of the real-time hydrometeorological data and historical operation data of the small hydropower station group in the basin specifically includes: collecting water level, flow rate, rainfall, and temperature data of upstream and downstream basin power stations with a sampling frequency of 15 minutes / time; obtaining the operation status data of key equipment such as generator sets, transformers, and gates of the small hydropower station group with a sampling frequency of 1 minute / time; collecting water quality parameters such as dissolved oxygen, pH value, and turbidity of water quality monitoring sections with a sampling frequency of 1 hour / time; integrating data on historical flood events, equipment failure records, and ecological pollution events for at least 5 years.

[0021] In practical applications, water level and flow rate data can be obtained through the automatic collection system of river hydrological stations, and rainfall and temperature data can be obtained through meteorological stations. The operation status of generator sets can be obtained through the power station automation system, including parameters such as rotational speed, power, and vibration; the transformer status includes oil temperature, winding temperature, etc.; the gate status includes opening degree, working status, etc. Water quality monitoring data can be obtained through fixed monitoring stations or mobile monitoring devices. Historical data can be integrated and obtained from the databases of power station management systems, basin management departments, and environmental protection departments.

[0022] In a certain embodiment, for 15 small hydropower stations distributed on a tributary of the Yangtze River Basin, water level and flow data of 3 upstream hydrological stations and 4 downstream hydrological stations were collected, with a sampling frequency of 15 minutes per time; operation status data of 35 units of 15 power stations were collected, including parameters such as rotational speed, power, and vibration, with a sampling frequency of 1 minute per time; water quality parameters such as dissolved oxygen, pH value, and turbidity of 5 water quality monitoring sections were collected, with a sampling frequency of 1 hour per time; historical flood events, equipment failure records, and ecological pollution event data from 2017 to 2022 were integrated.

[0023] Preferably, the construction of the complex network of virtual small hydropower station groups specifically includes: taking each small hydropower station in the basin as a node, constructing a node attribute vector, including geographical coordinates, installed capacity, regulation capacity, and equipment parameters; establishing a directed weighted edge based on the hydraulic connection of the river channel, and comprehensively considering the hydraulic transmission time, flow correlation, and geographical proximity for the edge weight; calibrating the hydraulic connection parameters and edge weights between nodes according to the measured hydrological data; using the community detection algorithm to identify subgroups of small hydropower stations with close connections to form a hierarchical network structure.

[0024] Specifically, the node attribute vector can be expressed as: , where: is the attribute vector of the first node, is the geographical coordinate of the node, is the installed capacity (megawatts, MW), is the regulation capacity (hours), is the equipment parameter vector (including the number of units, transformer capacity, etc.).

[0025] The weight of the directed weighted edge can be calculated by the following formula: , where: is the edge weight from node i to node j (dimensionless), is the hydraulic transmission time (hours), is the flow correlation (a dimensionless number between 0 and 1), is the geographical proximity (kilometers), are the weight coefficients of the corresponding factors, and . According to actual experience, usually set .

[0026] The hydraulic transmission time can be calculated by the river channel length and the average flow velocity: , where: is the river channel length from node i to node j (kilometers), is the average flow velocity (km / h). The average flow velocity can be calculated by the Manning formula or determined from measured data. In practical applications, for a river channel with a length of 10 km and an average flow velocity of 3 km / h, the hydraulic transit time is approximately 3.33 hours.

[0027] The flow correlation can be calculated by the Pearson correlation coefficient: , where: is the flow rate (m³ / s) of node i at time t, is the average flow rate (m³ / s) of node i, is the flow rate (m³ / s) of node j at time , is the average flow rate (m³ / s) of node j, is the time delay (hours) of flow transmission, n is the number of samples (pieces), and t is the discrete time point in the time series. In practical applications, the flow correlation is usually between 0.6 - 0.9, and node pairs with a correlation greater than 0.7 are considered to have a significant hydraulic connection.

[0028] The geographical proximity can be calculated by the Euclidean distance: , To keep the edge weights within a reasonable range, usually , , are normalized.

[0029] The community detection algorithm uses the Louvain algorithm to identify subgroups of small hydropower stations by maximizing the modularity. The formula for calculating the modularity Q is as follows: , where: is the edge weight (dimensionless) between nodes i and i, is the degree (dimensionless) of node i, is the sum of the weights of all edges in the network (dimensionless), is the community to which node i belongs (dimensionless integer identifier), is the Kronecker function, which is 1 when and 0 otherwise, and are node indices.

[0030] In an embodiment of the present invention, a virtual complex network is constructed for 15 small hydropower stations in a tributary of the Yangtze River Basin. Through the Louvain algorithm, 3 communities are identified, corresponding to the upstream, middle, and downstream regions respectively, and the modularity Q value reaches 0.68, indicating that the network has obvious community structure characteristics.

[0031] Preferably, the construction of the hierarchical embedded risk transfer prediction architecture including a physical connection layer, a risk mapping layer, and a multi-domain coupling layer specifically includes: in the physical connection layer, establishing node connection relationships based on hydraulic characteristics to capture the transfer characteristics of water flow dynamics; in the risk mapping layer, constructing a risk measurement space, defining a risk state vector and a risk distance function to characterize the spatio-temporal heterogeneity of risks; in the multi-domain coupling layer, establishing a two-way coupling mechanism for flood - equipment failure, equipment failure - ecological pollution, and ecological pollution - flood risks to depict the interaction relationships among various types of risks.

[0032] As Figure 2 shown, the hierarchical embedded risk transfer prediction architecture of the present invention includes three closely coupled layers: a physical connection layer, a risk mapping layer, and a multi-domain coupling layer.

[0033] In the physical connection layer, node connection relationships are mainly established based on hydraulic characteristics. The transfer characteristics of water flow dynamics are mainly described by the Saint-Venant equations: , , where: is the cross-sectional area of the river channel (square meters), is the flow rate (cubic meters per second), is the time (seconds), is the distance along the river channel (meters), is the lateral inflow per unit length (cubic meters per second per meter), is the acceleration due to gravity (9.8 meters per square second), is the integral of water pressure in the x direction (cubic meters), is the riverbed slope (dimensionless), is the friction slope (dimensionless), and respectively represent partial derivatives with respect to time and space.

[0034] In practical applications, appropriate simplified models, such as the kinematic wave model or the diffusion wave model, can be selected according to the river channel characteristics to improve the calculation efficiency.

[0035] In the risk mapping layer, a risk measurement space is constructed to map the observed data in the physical layer to the risk state space. The risk state vector is defined as: , Wherein: is the risk state vector at time t (dimensionless), is the flood risk (a dimensionless number between 0 and 1), is the equipment failure risk (a dimensionless number between 0 and 1), is the ecological pollution risk (a dimensionless number between 0 and 1).

[0036] The risk distance function is used to characterize the spatio-temporal characteristics of risk transmission between different nodes and is defined as: , Wherein: is the distance (dimensionless) between the risk state of node at time t and the risk state of node j at time , is the risk transmission time (hours), is the weight coefficient of the th type of risk (dimensionless), , is the th type of risk value (a dimensionless number between 0 and 1) of node at time , is the th type of risk value (a dimensionless number between 0 and 1) of node at time, where k is the risk type index, taking values of 1, 2, and 3, corresponding to the three types of risks of flood, equipment failure, and ecological pollution respectively.

[0037] The risk spatio-temporal heterogeneity can be characterized by the following indicators: , , , Wherein: is the spatial heterogeneity index (dimensionless), is the proportion of the risk of node i in the total risk (dimensionless); is the time heterogeneity index (dimensionless), is the proportion of the risk at time t in the total risk (dimensionless); is the type heterogeneity index (dimensionless), is the proportion of the th type of risk in the total risk (dimensionless); is the total number of nodes in the network (pieces), is the natural logarithm function. These indicators are based on the Shannon entropy principle and are used to quantify the uneven distribution degree of risk in different dimensions.

[0038] In the multi-domain coupling layer, the coupling relationship between different types of risks is established and can be represented by the following coupling matrix: , Where: represents the influence coefficient of the i-th type of risk on the j-th type of risk (dimensionless), , corresponding to three types of risks: flood, equipment failure, and ecological pollution respectively. For example, represents the influence coefficient of flood on equipment failure, represents the influence coefficient of equipment failure on ecological pollution.

[0039] According to the analysis of measured data, in a typical small hydropower station group, the influence coefficient of flood on equipment failure is about 0.65 - 0.85, the influence coefficient of equipment failure on ecological pollution is about 0.50 - 0.70, and the feedback influence coefficient of ecological pollution on flood is about 0.20 - 0.30.

[0040] In an embodiment of the present invention, by analyzing the historical data of a small hydropower station group in a tributary of the Yangtze River Basin, the obtained coupling matrix is: , This matrix indicates that flood has a strong influence on equipment failure (0.78), the influence of equipment failure on ecological pollution is also significant (0.63), while the feedback influence of ecological pollution on flood and equipment failure is relatively weak (0.22 and 0.18).

[0041] Preferably, the processing of multi-source heterogeneous data by using the topology-aware input mechanism specifically includes: realizing multi-source data fusion, unifying the data formats and time tags of different sources and different sampling frequencies; extracting local topology features, medium-distance topology features, and global topology features to form a multi-scale topology feature representation; identifying short-term fluctuation features, medium-term trend features, and long-term periodic features to construct multi-scale time series features; selectively focusing on the feature changes of key nodes and key time periods through the attention mechanism.

[0042] In the multi-source data fusion link, it is necessary to solve the problem of inconsistent sampling frequencies of different data sources. The present invention uses the time interpolation method for data alignment to ensure that all data has a unified time tag. The specific method is: , Where: is the data value of the moment t that needs to be interpolated (consistent with the unit of the original data), and are the two nearest sampling moments (hours) before and after t respectively, and are the measured data values (in the same unit as the original data) at times and respectively. is the time point (in hours) for which interpolation is required, is the time weight factor (dimensionless, ranging from 0 to 1). This linear interpolation method is applicable to parameters with relatively smooth changes such as water level and flow rate. For data with significantly different sampling frequencies, such as hydrological data (once every 15 minutes) and water quality data (once every hour), the system will first uniformly interpolate all data to the minimum time interval (15 minutes) and then perform subsequent processing to ensure the accuracy of time alignment.

[0043] The data standardization process uses the Z-score method: , where: is the standardized data (dimensionless), X is the original data (in the same unit as the original data), is the mean value (in the same unit as the original data), is the standard deviation (in the same unit as the original data).

[0044] Topological feature extraction is one of the core innovations of this invention. Local topological features mainly characterize the first-order neighborhood structure of nodes and can be calculated through the following indicators: , where: is the local topological feature vector of node i (dimensionless), is the degree of node i (dimensionless), is the clustering coefficient of node i (dimensionless), is the betweenness centrality of node i (dimensionless).

[0045] The medium-distance topological features mainly characterize the network structure within a 2 - 3 hop range: , where: is the medium-distance topological feature vector of node i (dimensionless), is the number of 2-hop neighbors of node i (in units), is the number of 3-hop neighbors of node i (in units), is the effective radius of node i (in hops).

[0046] The global topological features are mainly based on spectral graph theory and are extracted through the eigenvalues and eigenvectors of the Laplacian matrix: , , Wherein: is the Laplacian matrix (dimensionless), is the degree matrix (dimensionless), is the adjacency matrix (dimensionless), is the normalized Laplacian matrix (dimensionless), is the identity matrix (dimensionless), represents the diagonal matrix formed by taking the -1 / 2 power of the elements of the diagonal matrix D.

[0047] The global topological features can be represented as the eigenvectors corresponding to the first k smallest non-zero eigenvalues of the Laplacian matrix: , Wherein: is the global topological feature matrix (dimensionless), is the eigenvector corresponding to the i-th eigenvalue (dimensionless). In practical applications, usually k = 5 - 10, that is, the first 5 - 10 eigenvectors are extracted as the global topological features.

[0048] The extraction of time series features also adopts a multi-scale strategy. The short-term fluctuation features are mainly captured by moving window statistics: , Wherein: is the short-term fluctuation feature vector at time t (dimensionless), is the data within the window centered at time t (raw data), is the mean value of the data within the window, is the standard deviation of the data within the window, is the maximum value of the data within the window, is the minimum value of the data within the window, is the range of the data within the window. The window width is usually 6 - 24 hours.

[0049] The medium-term trend features are mainly extracted by the trend decomposition method: , Wherein: is the original time series (raw data), is the trend component (raw data), is the seasonal component (raw data), is the residual component (raw data). The medium-term trend features are .

[0050] The long-term periodic features are mainly extracted by Fourier transform or wavelet transform: , Wherein: is the Fourier transform of, representing the amplitude (complex number) of different frequency components, is the imaginary unit ( is the angular frequency (radians / second), is the base of the natural logarithm (approximately equal to 2.71828), represents the integration differential element with respect to time.

[0051] The attention mechanism is used to selectively focus on the feature changes of key nodes and key time periods, and can be expressed as: , , where: is the attention coefficient of node i to node i (a dimensionless number between 0 and 1), is the similarity coefficient between nodes i and j (dimensionless), is the similarity function, is the weight matrix (dimensionless), and are the feature vectors of nodes i and j respectively (dimensionless), is the exponential function with the base of the natural logarithm e, is the total number of nodes in the network (pieces).

[0052] In an embodiment of the present invention, the topological perception input mechanism is applied to the data of a small hydropower station group in a tributary of the Yangtze River Basin, and the local topological features, medium-distance topological features and global topological features of 15 nodes, as well as the short-term fluctuation features, medium-term trend features and long-term cycle features within 1 year are extracted. Through the attention mechanism, the system can automatically identify the feature changes of key nodes (such as large hydropower stations with strong regulation capabilities) and key time periods (such as flood seasons), improving the pertinence and accuracy of prediction.

[0053] Preferably, the capturing of the temporal dynamic characteristics through the risk state memory mechanism specifically includes: constructing multi-scale state representations of single-station state representation, local network state representation and global network state representation; achieving a dynamic balance between short-term memory update and long-term memory maintenance to retain key historical information; establishing a non-linear state transition mechanism based on threshold triggering to capture the sharp state changes caused by emergencies; and implementing smooth transition processing before and after the state transition to ensure the continuity of the prediction results.

[0054] The risk state memory mechanism is another core innovation point of the present invention, mainly based on the improved design of the long short-term memory network (LSTM). The multi-scale state representations include: 1. Single-station state representation: representing the risk state of a single hydropower station, with a dimension of 3 (flood, equipment failure, ecological pollution); 2. Local network state representation: It represents the overall state of the small hydropower sub-network, with a dimension of 3×K, where K is the number of nodes in the sub-network; 3. Global network state representation: It represents the system state of the small hydropower plant group in the entire basin, with a dimension of 3×N, where N is the total number of nodes in the network.

[0055] The dynamic balance between short-term memory update and long-term memory maintenance is achieved through the gating mechanism of LSTM. The specific formula is as follows: , , , , , , Among them: is the forget gate, which controls the retention ratio of the previous moment's memory (a vector between 0 and 1); is the input gate, which controls the reception ratio of the new input information (a vector between 0 and 1); is the candidate memory unit (vector); is the current memory unit (vector); is the output gate, which controls the output ratio of the current memory (a vector between 0 and 1); is the hidden state (vector); is the input at the current moment (vector); is the corresponding weight matrix (dimensionless); is the corresponding bias vector (dimensionless); is the sigmoid activation function, defined as , and the output range is (0, 1); is the hyperbolic tangent activation function, defined as , and the output range is (-1, 1); is the Hadamard product (element-wise multiplication); represents concatenating the vectors and column-wise into a new vector.

[0056] The present invention improves the standard LSTM by adding a non-linear state transition mechanism to capture the sharp state changes caused by sudden events: , , Among them: is the transition indicator (a dimensionless number with a value of 0 or 1), is the Heaviside step function (outputs 1 when the input is greater than 0, otherwise outputs 0), is the state cumulative value (dimensionless), is the transition threshold (dimensionless), is the original candidate memory unit (vector, obtained by standard LSTM calculation), is the updated candidate memory unit (vector, as the final used), is the candidate memory unit after transition (vector, representing the memory unit value in the burst state). The core idea of the state transition mechanism is: when the system state cumulative value triggers the transition mechanism, and uses the transition state to replace the original state ; otherwise, keep the original state unchanged. can be generated by a sub-model specifically trained for emergencies, or obtained by adding a preset transition vector to the original state: , where: is the transition vector (vector, representing the additional memory change in the burst state), learned from historical emergency data.

[0057] In practical applications, based on historical data analysis, an appropriate transition threshold can be set. For example, for flood risk, the flow rate exceeding 1.2 times the flood flow rate with a return period of 10 years can be set as the transition threshold; for equipment failure risk, the equipment abnormal index exceeding 1.5 times the safety value can be set as the transition threshold; for ecological pollution risk, the water quality parameter exceeding 1.3 times the standard limit can be set as the transition threshold.

[0058] To ensure the continuity of the prediction results, the present invention performs a smooth transition process before and after the state transition: , where: is the smoothed memory unit (vector), is the original memory unit (vector), is the predicted memory unit (vector), is the smoothing coefficient, ranging from [0, 1] (dimensionless). The smoothing coefficient can be dynamically adjusted over time, gradually increasing before the transition and gradually decreasing after the transition to achieve a smooth transition.

[0059] In one embodiment of the present invention, the risk state memory mechanism is applied to a small hydropower station group in a tributary of the Yangtze River Basin, and a flood risk transition event triggered by sudden heavy rainfall upstream is successfully captured. The system issues a warning 4.5 hours before the flood arrives, about 2 hours earlier than the traditional method, winning valuable time for the flood control and operation of downstream power stations. At the same time, through smooth transition processing, the prediction results maintain good continuity, avoiding frequent fluctuations of warning signals and improving the reliability of early warnings.

[0060] Preferably, generating the risk transfer prediction result through the multi-domain coupling output mechanism specifically includes: decoding the internal state into a flood risk level, equipment failure probability, and ecological pollution index respectively; generating short-term risk predictions for the next 24 hours and medium- and long-term risk trends for the next 7 days; predicting the time sequence, transfer intensity, and coupling effect of risk transfer between different sites; and determining the warning level, warning range, and suggestions for response measures based on the prediction results.

[0061] The multi-domain coupling output mechanism is a key link to realize the prediction of risk chain transfer. First, the internal state needs to be decoded into various risk indicators:

[0062] Decoding of flood risk level: , Where: is the probability distribution of the flood risk level (a 5-dimensional vector corresponding to the probabilities of 5 risk levels), usually divided into 5 levels (no risk, minor, medium, severe, extreme); and are the corresponding weight matrix and bias vector (dimensionless); is the hidden state (vector) of the LSTM; Softmax is the softmax function, defined as Softmax( ) = , where is the i-th element of the vector , K is the dimension of the vector , and this function converts the input into a probability distribution so that the sum of all elements is 1.

[0063] Decoding of equipment failure probability: , Where: is the equipment failure probability (a dimensionless number between 0 and 1), and are the corresponding weight matrix and bias vector (dimensionless); Sigmoid is the sigmoid function, defined as Sigmoid( ) = , and the output range is (0, 1).

[0064] Ecological pollution index decoding: , where: is the ecological pollution index vector (multiple water quality parameter values), which can include various water quality parameters (such as dissolved oxygen, pH value, turbidity, etc.); and are the corresponding weight matrix and bias vector (dimensionless).

[0065] Short-term risk prediction (24 hours) and medium- and long-term risk trend prediction (7 days) adopt a recursive prediction method: , , where: is the hidden state (vector) at the next moment, is the predicted input (vector) at the next moment, is the input prediction function, and LSTM represents the calculation process of the long short-term memory network defined above, is the input (vector) at the current moment. For longer-term predictions, a multi-step recursive strategy is adopted, that is, the prediction result of the previous step is used as the input for the next step to continue the prediction.

[0066] Input prediction function The specific construction method of , where: and are the weight matrices of the first and second layers (dimensionless), respectively, and are the bias vectors of the first and second layers (dimensionless), respectively. ReLU is the rectified linear unit activation function, defined as ReLU represents concatenating the vectors and by columns into a new vector.

[0067] This construction method enables the system to predict the input at the next moment based on the input and hidden state at the current moment, thereby realizing multi-step recursive prediction. In practical applications, to reduce error accumulation, actual observation data is introduced for correction after a certain number of steps: , where: is the predicted input (vector) at time t + n, is the actual observation data (vector) at time t + n, is the hidden state (vector) at time t + n - 1. The calibration interval is usually set to 3 to 6 time steps and adjusted according to data characteristics and prediction requirements. This method combining recursive prediction and regular calibration can avoid the problem of error accumulation in long-term prediction while maintaining the continuity of prediction, and is especially suitable for systems with complex dynamic characteristics such as small hydropower station groups in river basins.

[0068] Risk transfer prediction is the core output of the present invention, including transfer time, transfer intensity, and coupling effect: Transfer time prediction: , where: is the predicted time (hours) for the risk to transfer from node i to node j, is the reference transfer time (hours) under normal hydrological conditions, is the current flow rate (cubic meters per second) of node i, is the reference flow rate (cubic meters per second) of node i, is the adjustment coefficient (dimensionless), usually taking values of 0.2 - 0.5.

[0069] Transfer intensity prediction: , where: is the intensity of the risk transferring from node i to node j (a dimensionless number between 0 and 1), is the risk intensity of node i (a dimensionless number between 0 and 1), is the distance (kilometers) between nodes i and j, is the attenuation coefficient (1 / kilometer), is the coupling coefficient (dimensionless) between risk types k and l, is the coupling adjustment coefficient (dimensionless), is the base of the natural logarithm (approximately equal to 2.71828). Among them, k and l are the type indices of the risk source and the target risk respectively, taking values of 1, 2, 3, corresponding to three types of risks: flood, equipment failure, and ecological pollution.

[0070] Coupling effect prediction: , where: is the coupling effect (dimensionless) between nodes i and j, is the coupling weight (dimensionless) between risk types k and l, is the k-th type of risk of node i (a dimensionless number between 0 and 1), is the l-th type of risk of node j (a dimensionless number between 0 and 1). The coupling effect describes the interaction intensity between different types of risks.

[0071] Based on the prediction results, the early warning level, scope, and suggestions for response measures can be determined. The early warning level is usually divided into four levels: blue (general), yellow (relatively serious), orange (severe), and red (extremely serious). The early warning scope is determined according to the risk transmission path, and the suggestions for response measures are formulated based on the risk type and level.

[0072] In one embodiment of the present invention, the system conducts risk chain transmission prediction for a group of small hydropower stations in a tributary of the Yangtze River Basin, and successfully predicts a risk transmission process starting from an upstream power station and spreading to a downstream power station through a midstream power station. The prediction shows that the flood risk of the upstream power station will be transmitted to the midstream power station after 6 hours, increasing the equipment failure risk of the midstream power station from 0.15 to 0.68, and then being transmitted to the downstream power station after 4 hours, increasing the ecological pollution risk of the downstream power station from 0.08 to 0.52. Based on this, the system issues a yellow early warning, arranges equipment maintenance for the midstream power station and water quality monitoring for the downstream power station in advance, and successfully avoids a potential equipment failure and water quality pollution incident.

[0073] In another embodiment of the present invention, it also includes the verification and optimization process of the risk chain transmission model: verifying the fitting ability of the model through the reproduction of historical events; analyzing the sources of prediction errors and biases by comparing actual observations with predictions; optimizing the model parameters and structure based on the verification results; establishing a prediction accuracy evaluation mechanism, and the prediction accuracy reaches more than 90%.

[0074] Model verification and optimization are the key links to ensure the reliability of the prediction system. The reproduction of historical events uses the backtesting method, selects typical historical events for simulation prediction, compares the prediction results with actual observations, and calculates the goodness of fit: , where: is the goodness of fit (dimensionless), is the actual observed value (original unit), is the predicted value (original unit), is the mean of the observed values (original unit), is the number of samples (pieces). Ideally, should be close to 1, indicating that the model can fit the actual data well. The prediction error analysis uses the root mean square error (RMSE) and the mean absolute percentage error (MAPE): , , where: RMSE is the root mean square error (consistent with the original data unit), and MAPE is the mean absolute percentage error (%).

[0075] According to the error analysis results, the gradient descent method is used to optimize the model parameters: , where: is the current parameter (dimensionless), is the updated parameter (dimensionless), is the learning rate (dimensionless), is the loss function with respect to the parameter gradient (dimensionless).

[0076] The prediction accuracy evaluation is calculated based on the confusion matrix: , where: is the prediction accuracy (%), is the number of true positives (pieces), is the number of true negatives (pieces), is the number of false positives (pieces), is the number of false negatives (pieces).

[0077] In the embodiment of the present invention, the risk chain transmission prediction system of a small hydropower station group in a tributary of the Yangtze River Basin is verified, and 15 typical risk events between 2017 and 2022 are selected for backtesting. The results show that the goodness of fit R 2 of the model for flood risk reaches 0.92, the goodness of fit for equipment failure risk reaches 0.88, and the goodness of fit for ecological pollution risk reaches 0.85. In terms of prediction accuracy, the 24-hour short-term prediction accuracy reaches 94.3%, and the 7-day medium- and long-term prediction accuracy reaches 91.5%, meeting the requirements of practical applications.

[0078] Preferably, the application scenarios of the risk chain transmission prediction method for the small hydropower station group in the basin include: flood risk warning during the flood season, discovering potential risks 4 - 6 hours in advance; equipment failure prediction and preventive maintenance, reducing the failure rate by more than 30%; ecological environmental pollution warning and formulation of protection measures, reducing the incidence of ecological pollution events by more than 50%; coordinated scheduling optimization of the small hydropower station group in the basin, improving the water resource utilization efficiency by more than 15%.

[0079] The risk chain transmission prediction system of the small hydropower station group in the basin proposed by the present invention is used to implement the above method, including: a data acquisition module, a network construction module, a risk mapping module, a multi-domain coupling module, a prediction analysis module, a decision support module, and a model optimization module.

[0080] The data acquisition module is used to obtain the real-time hydrometeorological data and historical operation data of the small hydropower station group in the basin. Specifically, it includes a hydrological data acquisition unit, an equipment status acquisition unit, a water quality monitoring unit, and a historical data management unit. Among them, the hydrological data acquisition unit is responsible for collecting the water level, flow rate, rainfall, and temperature data of the upstream and downstream basin power stations; the equipment status acquisition unit is responsible for obtaining the operation status data of key equipment such as generator sets, transformers, and gates in the small hydropower station group; the water quality monitoring unit is responsible for collecting water quality parameters such as dissolved oxygen, pH value, and turbidity at the water quality monitoring section; the historical data management unit is responsible for integrating historical flood events, equipment failure records, and ecological pollution event data.

[0081] The network construction module is used to construct a complex network of virtual small hydropower station groups. Specifically, it includes a node construction unit, an edge establishment unit, a weight calculation unit, and a community identification unit. Among them, the node construction unit is responsible for constructing a node attribute vector with each small hydropower station in the basin as a node; the edge establishment unit is responsible for establishing a directed weighted edge based on the hydraulic connection of the river; the weight calculation unit is responsible for calculating the edge weight, comprehensively considering the hydraulic transmission time, flow correlation degree, and geographical proximity; the community identification unit is responsible for identifying subgroups of small hydropower stations with close connections through community detection algorithms.

[0082] The risk mapping module is used to realize the spatio-temporal heterogeneity characterization of risks. Specifically, it includes a risk measurement unit, a risk distance calculation unit, and a risk heterogeneity analysis unit. Among them, the risk measurement unit is responsible for constructing a risk measurement space and defining a risk state vector; the risk distance calculation unit is responsible for calculating the spatio-temporal characteristics of risk transmission between different nodes; the risk heterogeneity analysis unit is responsible for characterizing the heterogeneity of risks in space, time, and type.

[0083] The multi-domain coupling module is used to establish the coupling relationship between floods - equipment failures - ecological pollution. Specifically, it includes a flood - equipment coupling unit, an equipment - ecology coupling unit, and an ecology - flood coupling unit. Among them, the flood - equipment coupling unit is responsible for establishing the influence mechanism of floods on equipment failures; the equipment - ecology coupling unit is responsible for establishing the influence mechanism of equipment failures on ecological pollution; the ecology - flood coupling unit is responsible for establishing the feedback influence mechanism of ecological pollution on floods.

[0084] The prediction and analysis module is used to generate risk transmission prediction results using an embedded time series deep learning framework. Specifically, it includes a data preprocessing unit, a topological feature extraction unit, a time series feature extraction unit, an LSTM prediction unit, and a result decoding unit. Among them, the data preprocessing unit is responsible for data cleaning, standardization, and alignment; the topological feature extraction unit is responsible for extracting multi-scale topological features; the time series feature extraction unit is responsible for extracting multi-scale time series features; the LSTM prediction unit is responsible for generating predictions based on the risk state memory mechanism; the result decoding unit is responsible for decoding the prediction results into flood risk levels, equipment failure probabilities, and ecological pollution indicators.

[0085] The decision support module is used to generate early warning information and intervention measure suggestions based on the prediction results. Specifically, it includes an early warning generation unit, a scope determination unit, and a measure suggestion unit. Among them, the early warning generation unit is responsible for generating hierarchical early warning information based on the prediction results; the scope determination unit is responsible for determining the early warning scope and objects; the measure suggestion unit is responsible for generating risk prevention and control suggestion measures.

[0086] The model optimization module is used to implement model verification and parameter optimization. Specifically, it includes a verification unit, an error analysis unit, and a parameter optimization unit. Among them, the verification unit is responsible for verifying the fitting ability of the model through the reproduction of historical events; the error analysis unit is responsible for analyzing the sources of prediction errors and deviations; the parameter optimization unit is responsible for optimizing the model parameters and structure based on the verification results.

[0087] During the actual operation of the system, there is a close data flow relationship among the modules. The data acquisition module first obtains real-time data and historical data, and simultaneously transmits the processed data to the network construction module and the prediction analysis module. The network construction module constructs a virtual complex network based on the received data and transmits the network structure information to the risk mapping module. The risk mapping module realizes the mapping from the physical space to the risk space and transmits the risk characterization information to the multi-domain coupling module. The multi-domain coupling module establishes the coupling relationship of different risk types and transmits the coupling matrix to the prediction analysis module. The prediction analysis module integrates all the information, generates the risk transmission prediction results, and transmits the results to the decision support module and the model optimization module. The decision support module generates early warning information and intervention measure suggestions based on the prediction results and provides them to the user. The model optimization module, based on the difference between the prediction results and the actual observations, feeds back the optimized parameters to the prediction analysis module to form a closed-loop optimization.

[0088] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, various changes and modifications can be made to the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. Risk chain transfer prediction method for small hydropower station groups in a basin, characterized in that Including: Obtain the real-time hydrometeorological data and historical operation data of the small hydropower station group in the basin; including: Based on the real-time hydrometeorological data and historical operation data, construct a hierarchical embedded risk transfer prediction architecture including a physical connection layer, a risk mapping layer, and a multi-domain coupling layer; According to the hierarchical embedded risk transfer prediction architecture, construct a complex network of virtual small hydropower station groups; Based on the complex network of virtual small hydropower station groups, establish a multi-domain coupling model of flood - equipment failure - ecological pollution using an embedded time series deep learning framework; including: Use a topology-aware input mechanism to process multi-source heterogeneous data; Capture the time series dynamic characteristics through a risk state memory mechanism; Generate risk transfer prediction results through a multi-domain coupling output mechanism; Output the risk transfer prediction results to achieve the prediction of risk chain transfer in the small hydropower station group in the basin.

2. The method according to claim 1, wherein The specific process of obtaining the real-time hydrometeorological data and historical operation data of the small hydropower station group in the basin includes: Collect the water level, flow rate, rainfall, and temperature data of the upstream and downstream basin power stations, with a sampling frequency of 15 minutes / time; Obtain the operation status data of key equipment such as generator sets, transformers, and gates of the small hydropower station group, with a sampling frequency of 1 minute / time; Collect water quality parameters such as dissolved oxygen, pH value, and turbidity of the water quality monitoring section, with a sampling frequency of 1 hour / time; Integrate historical flood event, equipment failure record, and ecological pollution event data for at least 5 years.

3. The method according to claim 1, wherein The specific process of constructing a complex network of virtual small hydropower station groups includes: Taking each small hydropower station in the basin as a node, construct a node attribute vector, including geographical coordinates, installed capacity, regulation capacity, and equipment parameters; Establish a directed weighted edge based on the hydraulic connection of the river channel, and comprehensively consider the hydraulic transfer time, flow correlation, and geographical proximity for the edge weight; Calibrate the hydraulic connection parameters and edge weights between nodes according to the measured hydrological data; Use the community detection algorithm to identify subgroups of small hydropower stations with close connections to form a hierarchical network structure.

4. The method according to claim 1, wherein The specific process of constructing a hierarchical embedded risk transfer prediction architecture including a physical connection layer, a risk mapping layer, and a multi-domain coupling layer includes: In the physical connection layer, establish a node connection relationship based on hydraulic characteristics to capture the water flow dynamics transfer characteristics; In the risk mapping layer, construct a risk measurement space, define a risk state vector and a risk distance function to characterize the spatio-temporal heterogeneity of risks; In the multi-domain coupling layer, establish a two-way coupling mechanism for flood - equipment failure, equipment failure - ecological pollution, and ecological pollution - flood risks to depict the interaction relationship between multiple types of risks.

5. The method according to claim 1, wherein The specific process of using a topology-aware input mechanism to process multi-source heterogeneous data includes: Realize multi-source data fusion to unify the data formats and time tags of different sources and different sampling frequencies; Extract local topology features, medium-distance topology features, and global topology features to form a multi-scale topology feature representation; Identify short-term fluctuation features, medium-term trend features, and long-term cycle features to construct multi-scale time series features; Selectively focus on the feature changes of key nodes and key time periods through the attention mechanism.

6. The method according to claim 1, wherein The specific process of capturing the time series dynamic characteristics through a risk state memory mechanism includes: Construct a multi-scale state representation of single-station state representation, local network state representation, and global network state representation; Achieve the dynamic balance between short-term memory update and long-term memory maintenance, and retain key historical information; Establish a non-linear state transition mechanism triggered by thresholds to capture the sharp state changes caused by unexpected events; Implement smooth transition processing before and after state transitions to ensure the continuity of prediction results.

7. The method according to claim 1, characterized in that The generation of risk transfer prediction results through the multi-domain coupling output mechanism specifically includes: Decode the internal state into flood risk levels, equipment failure probabilities, and ecological pollution indicators respectively; Generate short-term risk predictions for the next 24 hours and medium- and long-term risk trend predictions for the next 7 days; Predict the time sequence, transfer intensity, and coupling effects of risk transfer between different sites; Determine the warning level, warning scope, and suggestions for countermeasures based on the prediction results.

8. The method according to claim 7, characterized in that, It also includes the verification and optimization process of the risk chain transfer model: Verify the fitting ability of the model through the reproduction of historical events; Use the comparison between actual observations and predictions to analyze the sources of prediction errors and biases; Optimize the model parameters and structure based on the verification results; Establish a prediction accuracy evaluation mechanism.

9. The method according to claim 8, wherein The application scenarios of the risk chain transfer prediction method for small hydropower station groups in the basin include: Flood risk warning during the flood season; Equipment failure prediction and preventive maintenance; Ecological environmental pollution warning and formulation of protection measures; Collaborative scheduling optimization of small hydropower station groups in the basin.

10. A risk chain transmission prediction system for small hydropower station groups in a river basin implementing the method according to any one of claims 1-9, characterized in that, It includes: A data acquisition module for obtaining real-time hydrometeorological data and historical operation data of small hydropower station groups in the basin; A network construction module for constructing a complex network of virtual small hydropower station groups; A risk mapping module for realizing the characterization of risk spatio-temporal heterogeneity; A multi-domain coupling module for establishing the coupling relationship between flood - equipment failure - ecological pollution; A prediction analysis module for generating risk transfer prediction results using an embedded time series deep learning framework; A decision support module for generating warning information and suggestions for intervention measures based on the prediction results; A model optimization module for realizing model verification and parameter optimization.

Images