Modularized integrated storage and drainage intelligent regulation and control construction method based on sponge city concept

Abstract

The application relates to the technical field of rainwater management and intelligent construction, and discloses a modularized storage and drainage integrated intelligent regulation and control construction method based on a sponge city concept. The method generates an initial cooperative instruction sequence by establishing a multilayer correlation network of mapping environmental disturbance factors, inputting a reference rainfall intensity, and activating a seepage state transition function layer by layer, and the generated initial cooperative instruction sequence is fed to physical units and feedback is collected; a state deviation topology graph is constructed based on the feedback and target differences, network node parameters are reversely adjusted to update the function and generate correction instructions; the process is iterated until the convergence condition is met, the final network parameters and instruction sequence are locked, and pre-configuration before construction is completed. The method realizes dynamic cooperative simulation of multiple environmental factors and closed-loop adaptive optimization of the model based on physical feedback, and improves the accuracy of system pre-configuration and the overall operation efficiency.

Description

Technical Field

[0001] This invention relates to the field of rainwater management and intelligent construction technology, specifically a modular integrated intelligent control construction method for water storage and drainage based on the concept of sponge cities. Background Technology

[0002] In existing sponge city construction, the construction and configuration of modular water storage and drainage facilities largely rely on static hydrological calculations and separate designs. A common practice is to determine the capacity of water storage modules, the distribution of infiltration facilities, and the size of the drainage network separately using empirical formulas based on historical rainfall data or design rainfall patterns, followed by assembly and construction. This technical approach simplifies and fragments the complex process of "infiltration, retention, storage, purification, utilization, and drainage," and its pre-construction simulations fail to fully consider the dynamic correlations and synergistic effects among various environmental disturbance factors. On-site construction configuration largely depends on the experience and judgment of designers, lacking a pre-simulation mechanism that can comprehensively simulate the interactions of multiple factors in virtual space and generate integrated, coordinated operational commands.

[0003] In the pre-construction configuration phase, existing technologies lack an effective closed-loop correction channel between the model and the physical entity. After initial design parameters are issued to the physical module, deviations between its actual operational performance and design expectations are typically only discovered after construction through monitoring, requiring subsequent manual adjustments or engineering modifications. This process fails to calibrate and optimize the digital model itself before construction, resulting in insufficient matching between theoretical design and actual site conditions, making it difficult to achieve optimal pre-configuration of system operational performance. A method is needed to achieve a closed loop from multi-factor dynamic collaborative deduction to model self-correction based on physical feedback, thereby locking in the optimal control strategy highly aligned with the site conditions before construction. Summary of the Invention

[0004] The purpose of this invention is to provide a modular, integrated, intelligent control construction method for water storage and drainage based on the concept of sponge cities, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a modular, integrated intelligent control construction method for water storage and drainage based on the concept of sponge cities, the method comprising:

[0006] A multi-layered correlation network of modular storage and discharge units is established, where each layer corresponds to an environmental disturbance factor; for the multi-layered correlation network, a set of benchmark rainfall intensities is introduced as the input source of the environmental disturbance factor.

[0007] Perform inter-layer collaborative inference operation: Based on the benchmark rainfall intensity, activate the seepage state transition function corresponding to each environmental disturbance factor in the multi-layer interconnected network layer by layer to generate an initial collaborative seepage command sequence;

[0008] Implement dynamic mapping and response of field units: send the initial coordinated seepage command sequence to the corresponding modular storage and discharge unit, and at the same time collect the operation status feedback information of the modular storage and discharge unit in the physical space;

[0009] The simulation and feedback are used for iterative correction: Based on the operational status feedback information and the target status of the benchmark rainfall intensity, a state deviation topology is constructed. The state deviation topology is used to adjust the node parameters of the corresponding environmental disturbance factor layer in the multi-layer interconnected network in reverse order to update the seepage state transition function and generate a corrected coordinated seepage command sequence.

[0010] The dynamic mapping and response steps and subsequent cyclic correction steps are executed iteratively until the running status feedback information meets the convergence condition under the benchmark rainfall intensity. The current network parameters and the final coordinated seepage command sequence are then locked to complete the pre-configuration before construction.

[0011] Preferably, the establishment of the multi-layer interconnected network of the modular storage and discharge unit specifically includes:

[0012] Identify a set of environmental disturbance factors that affect the performance of the modular storage and discharge unit, including rainfall intensity, surface material, groundwater level, and vegetation cover.

[0013] Each environmental disturbance factor in the set of environmental disturbance factors is assigned an independent network layer, and within each network layer, a topological node reflecting the hydraulic connection between the units is constructed based on the geographical layout relationship of the modular storage and discharge units.

[0014] Between adjacent network layers, cross-layer connection edges are established based on the coupling relationship between environmental disturbance factors. These cross-layer connection edges are used to transmit the superimposed influence signal of different environmental disturbance factors on the seepage process.

[0015] Each topology node is configured with a trainable set of node parameters, and an intra-layer aggregation rule is defined for each network layer to integrate input signals from lower-layer networks or cross-layer connection edges to form the output features of the corresponding network layer.

[0016] Preferably, the execution of the inter-layer collaborative inference operation specifically includes:

[0017] The reference rainfall intensity is input into the network layer corresponding to the rainfall intensity as the initial excitation signal of the rainfall intensity layer;

[0018] According to the preset order from the rainfall intensity layer to the vegetation coverage layer, the seepage state transfer function is activated layer by layer; for the current layer, the seepage state transfer function takes the feature vector output by the previous layer and the correlation signal passed through the cross-layer connection edge as input, and calculates it in combination with the node parameter set of the current layer to output the seepage state vector of the current layer.

[0019] During the calculation process of each layer, an intra-layer aggregation operation is performed simultaneously. Based on the intra-layer aggregation rules, the output states of all topological nodes in the current layer are integrated to generate the aggregated state characteristics of the current network.

[0020] The percolation state vector output by the last layer network is concatenated and formatted with the aggregated state features of all intermediate layers according to a preset encoding rule to generate the initial collaborative percolation instruction sequence.

[0021] Preferably, the specific steps of activating the seepage state transfer function layer by layer according to a preset order from the rainfall intensity layer to the vegetation cover layer include:

[0022] A fixed activation sequence of environmental disturbance factors is determined, which starts at the rainfall intensity layer and ends at the vegetation cover layer;

[0023] The reference rainfall intensity is input to the rainfall intensity layer as the initial input excitation signal;

[0024] For the current activated layer in the sequence, the output feature vector from the previous activated layer in the sequence is received as the first input, and the correlation signal from other related layers through all defined cross-layer connection edges is received as the second input.

[0025] The first input and the second input are concatenated as vectors to form the complete input vector of the current layer.

[0026] The percolation state transition function configured in the current active layer is called to calculate the complete input vector and the node parameter set of the current active layer, and output the percolation state vector of the current layer.

[0027] The percolation state vector output from the current layer is passed to the next layer in the activation sequence as the first input to the next layer, until all layers in the activation sequence have been calculated.

[0028] Preferably, the step of simultaneously performing intra-layer aggregation during the calculation process of each layer, and integrating the output states of all topological nodes in the current layer according to the intra-layer aggregation rules to generate the aggregated state features of the current network, specifically includes:

[0029] After the seepage state transition function of the current layer completes the state calculation for all topology nodes, the output state vector of all topology nodes in the current layer is obtained.

[0030] Based on the predefined intra-layer aggregation rules for the current layer, the weighting coefficients of the output state vector of each topology node in the aggregation process are determined; wherein, the weighting coefficients are jointly determined by the centrality of the topology node in the geographic layout relationship and the matching degree of the current input signal.

[0031] According to the weighting coefficients, the output state vectors of all topological nodes in the current layer are weighted and summed to obtain a preliminary aggregate vector.

[0032] The initial aggregation vector is subjected to nonlinear transformation and dimensional reduction to generate the aggregated state features that represent the overall seepage state characteristics of the current layer.

[0033] Preferably, the dynamic mapping and response of the implementation site unit specifically includes:

[0034] The initial coordinated seepage command sequence is parsed and decoded into a set of control commands for each specific modular storage and discharge unit. The set of control commands includes target seepage rate, valve opening degree and pumping power information.

[0035] The control command set is synchronously sent to the corresponding modular storage and discharge unit at the construction site for execution through a distributed control system.

[0036] During the execution of instructions by the modular storage and discharge unit, physical operation data is collected in real time through sensor arrays deployed at each unit and connecting pipes. The physical operation data includes actual flow rate, water level and storage volume.

[0037] The physical operation data is bound to a preset timestamp and unit identifier, and after data fusion processing, structured operation status feedback information with spatial location is generated.

[0038] Preferably, the cyclical correction of deduction and feedback specifically includes:

[0039] By comparing the actual state of each modular storage and discharge unit in the operation status feedback information with the expected target state under the benchmark rainfall intensity, the state difference of each unit in each dimension is calculated.

[0040] Based on the topological connection relationship between units, the state difference of each unit is propagated along the connection path in the multi-layer association network to calculate the responsibility error that each topological node should bear.

[0041] Based on the distribution of the responsibility error, the state deviation topology diagram is constructed, which records the intensity and direction of the error at each layer and node of the network.

[0042] A gradient-based parameter adjustment strategy is adopted. Based on the state deviation topology graph, the node parameter set of the relevant network layer in the multi-layer interconnected network is iteratively fine-tuned from top to bottom. The direction of fine-tuning is to minimize the responsibility error.

[0043] After each fine-tuning, the seepage state transition function is run again to generate a new cooperative seepage command sequence as the corrected cooperative seepage command sequence.

[0044] Preferably, the specific steps for calculating the state difference of each unit in each dimension by comparing the actual state of each modular storage and discharge unit in the operational status feedback information with the expected target state under the benchmark rainfall intensity include:

[0045] Extract the actual status data corresponding to the identifier of each modular storage and discharge unit from the operation status feedback information. The actual status data includes the actual flow velocity value, the actual water level height value, and the actual water storage volume value.

[0046] Based on the baseline rainfall intensity, the target state data corresponding to each modular storage and drainage unit identifier is retrieved from the predefined target state mapping table. The target state data includes the target flow velocity value, the target water level height value, and the target water storage volume value.

[0047] For each modular storage and discharge unit, in terms of flow velocity, the absolute difference between the actual flow velocity value and the target flow velocity value is calculated as the flow velocity state difference quantity.

[0048] In the dimension of water level height, the absolute difference between the actual water level height value and the target water level height value is calculated as the water level state difference quantity;

[0049] In terms of water storage volume, the absolute difference between the actual water storage volume and the target water storage volume is calculated as the volume state difference.

[0050] Preferably, the specific steps for calculating the responsibility error to be borne by each topological node by propagating the state difference of each unit along the connection path in the multi-layered network based on the topological connection relationship between units include:

[0051] The state difference of each modular storage and drainage unit is mapped to the topological node of the corresponding geographical location in the multi-layer interconnected network, which serves as the initial error signal of the topological node.

[0052] Based on the node connection relationships defined in the multi-layered network, an error propagation directed graph is constructed, wherein the weights of the node connection edges are assigned according to the strength of the hydraulic connection.

[0053] Using an iterative propagation algorithm, in the directed error propagation graph, the initial error signal of each node is weighted and distributed to the forward nodes along its output connection edge, while the error distribution signal from the backward nodes is received along its input connection edge.

[0054] After multiple iterations until the error allocation stabilizes, the sum of all error allocation signals received by each topology node is calculated, and the sum of the received error allocation signals is added to the initial error signal of the topology node itself. The sum is taken as the responsibility error that the topology node should bear.

[0055] Preferably, the specific steps of adopting a gradient-based parameter adjustment strategy, which iteratively fine-tunes the node parameter set of relevant network layers in the multi-layer interconnected network from top to bottom according to the state deviation topology graph, with the fine-tuning direction being to minimize the responsibility error, include:

[0056] Extract the responsibility error value of each topology node and its gradient direction relative to each parameter in the node parameter set from the state deviation topology graph;

[0057] Starting from the vegetation cover layer, parameters are adjusted sequentially for each relevant network layer in reverse order; for the current adjustment layer, each topology node within the adjustment layer is traversed.

[0058] For each parameter in the node parameter set of the current topology node, a parameter update step size is calculated based on the gradient direction information corresponding to the parameter provided by the state deviation topology graph; the parameter update step size is directly proportional to the responsibility error value of the current node and inversely proportional to the average amplitude of historical parameter updates.

[0059] Update the step size according to the parameters, and adjust the value of the parameters in the opposite direction of the gradient direction;

[0060] After all parameters of all topology nodes in the current layer have been adjusted, the percolation state vector of the adjusted layer is recalculated using the updated node parameter set, and the overall responsibility error is verified to have decreased. If it has not decreased, the parameter update step size is reduced proportionally and the adjustment steps of this layer are re-executed until the overall responsibility error meets the minimization trend requirement, and then the parameters of the next layer network are adjusted.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] By establishing a multi-layered correlation network for various environmental disturbance factors and importing a benchmark rainfall intensity as a unified input source, the corresponding seepage state transition functions in each network layer are sequentially activated and collaboratively calculated. This process transforms the originally isolated influence of environmental factors into a networked dynamic interaction simulation, realizing an integrated dynamic extrapolation of the entire process of rainwater storage, retention, infiltration, and drainage under complex disturbance conditions. This generates a coordinated initial seepage command sequence, surpassing the limitations of conventional static design or single-factor models in reflecting system coupling effects.

[0063] By sending initial command sequences to physical modules and collecting their operational status feedback, a state deviation topology diagram representing the distribution and transmission path of deviations between the target and measured values ​​is constructed. This diagram is then used to adjust the parameters of corresponding nodes in the multi-layer network in reverse. This technology establishes a precise feedback and adaptive correction closed loop between the "digital model and physical entity," enabling the transfer function describing the seepage process to be iteratively optimized based on the actual response. Ultimately, this allows the digital model to infinitely approximate the real hydrological response characteristics of a specific site. The locked network parameters and command sequences achieve highly accurate pre-configuration before construction, overcoming the shortcomings of conventional technologies where the model is fixed and cannot adaptively match site conditions. Attached Figure Description

[0064] Figure 1 This is a schematic diagram illustrating the working principle of the modular integrated intelligent regulation and control construction method for water storage and drainage based on the concept of sponge cities as described in this invention.

[0065] Figure 2 A flowchart for establishing a multi-layered interconnected network;

[0066] Figure 3 A flowchart for performing intra-layer aggregation operations;

[0067] Figure 4 A comparison diagram of error propagation in the water storage and drainage units of a sponge city;

[0068] Figure 5 A normalized comparison chart of the multi-dimensional state differences of the water storage and drainage units in a sponge city. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Please see Figure 1This invention provides a modular integrated intelligent control construction method for water storage and drainage based on the concept of sponge cities. The method includes: establishing a multi-layered interconnected network of modular water storage and drainage units, where each layer corresponds to an environmental disturbance factor; importing a set of benchmark rainfall intensities as input sources for the environmental disturbance factors into the multi-layered interconnected network; performing inter-layer collaborative inference operations: based on the benchmark rainfall intensities, activating the seepage state transition functions corresponding to each environmental disturbance factor layer by layer in the multi-layered interconnected network to generate an initial collaborative seepage command sequence; and implementing dynamic mapping and response of the field units: issuing the initial collaborative seepage command sequence to the corresponding modular water storage and drainage units, while simultaneously collecting data from the modular water storage and drainage units. The system receives operational status feedback information from the block-type storage and drainage unit in physical space; it then performs cyclical correction through deduction and feedback: based on the operational status feedback information and the target state of the benchmark rainfall intensity, a state deviation topology diagram is constructed. This diagram is then used to reverse-adjust the node parameters of the corresponding environmental disturbance factor layer in the multi-layered interconnected network to update the seepage state transition function, generating a corrected coordinated seepage command sequence. The system iteratively executes the dynamic mapping and response steps and subsequent cyclical correction steps until the operational status feedback information meets the convergence condition under the benchmark rainfall intensity. Finally, it locks the current network parameters and the final coordinated seepage command sequence, completing the pre-configuration before construction.

[0071] In one embodiment of the present invention, see [reference] Figure 2The method identifies a set of environmental disturbance factors affecting the performance of modular storage and drainage units, including rainfall intensity, surface material, groundwater level, and vegetation cover. Each environmental disturbance factor is assigned an independent network layer. Within each network layer, topological nodes reflecting the hydraulic connections between units are constructed based on the geographical layout of the modular storage and drainage units. Between adjacent network layers, cross-layer connections are established based on the coupling relationships between environmental disturbance factors. These cross-layer connections are used to transmit the superimposed influence signals of different environmental disturbance factors on the seepage process. A trainable set of node parameters is configured for each topological node, and an intra-layer aggregation rule is defined for each network layer to integrate input signals from lower-level networks or cross-layer connections, forming the output features of the corresponding network layer. The baseline rainfall intensity is input into the network layer corresponding to the rainfall intensity as the initial excitation signal for the rainfall intensity layer. The seepage state transition function is activated layer by layer according to a preset order from the rainfall intensity layer to the vegetation cover layer. For the current layer, the seepage state transition function uses the feature vector output from the previous layer and the associated signal passed through the cross-layer connection edge as input, combined with the node parameter set of the current layer, to calculate and output the seepage state vector of the current layer. During the calculation of each layer, an intra-layer aggregation operation is performed simultaneously, integrating the output states of all topological nodes in the current layer according to the intra-layer aggregation rules to generate the aggregated state features of the current network. The seepage state vector output from the last layer and the aggregated state features of all intermediate layers are concatenated and formatted according to a preset encoding rule to generate the initial cooperative seepage instruction sequence.

[0072] In some embodiments, the example scenario is an urban drainage area containing ten modular storage and drainage units. The set of environmental disturbance factors includes four factors: rainfall intensity, surface material, groundwater level, and vegetation cover, corresponding to the establishment of a four-layer interconnected network. Each layer of the network contains ten topology nodes, each corresponding to a modular storage and drainage unit. The connections between topology nodes are defined according to the actual pipeline layout between the modular storage and drainage units, forming a hydraulic connection topology. Cross-layer connection edges are added according to the physical coupling relationship between environmental disturbance factors. For example, a connection edge is established between the rainfall intensity layer and the surface material layer to represent the superimposed influence signal of rainfall on the surface seepage process. The trainable node parameter set of each topology node is initialized with random values, and the intra-layer aggregation rule is defined as a weighted summation operation. The baseline rainfall intensity input is rainfall data of 50 mm per hour, serving as the initial excitation signal for the rainfall intensity layer. The seepage state transfer function is activated layer by layer, from the rainfall intensity layer to the vegetation cover layer. For the surface material layer, the seepage state transfer function receives the feature vector output from the rainfall intensity layer and the associated signal transmitted from the rainfall intensity layer through cross-layer connection edges. Combined with the node parameter set of the surface material layer, it calculates and outputs the seepage state vector of the surface material layer. Simultaneously, an intra-layer aggregation operation is performed on the surface material layer, integrating the output state vectors of ten topological nodes according to the intra-layer aggregation rules to generate the aggregated state features of the surface material layer. The groundwater level layer and the vegetation cover layer are processed sequentially. The seepage state vector output from the vegetation cover layer is concatenated with the aggregated state features of the rainfall intensity layer, surface material layer, and groundwater level layer, and formatted to generate the initial coordinated seepage command sequence.

[0073] Optionally, when constructing a multi-layer interconnected network, the number of topological nodes corresponding to each environmental disturbance factor in the set of environmental disturbance factors is consistent with the actual number of modular storage and drainage units. Each topological node's trainable node parameter set includes a weight matrix and a bias vector, used for linear transformation and nonlinear activation in the seepage state transition function. The weighting coefficients in the intra-layer aggregation rules are jointly determined by the centrality of the topological node in the geographical layout relationship and the matching degree of the current input signal. The centrality is calculated based on the degree centrality of the modular storage and drainage unit connection graph, and the matching degree is calculated using cosine similarity. The specific form of the seepage state transition function adopts a fully connected neural network layer, whose input is a concatenated vector of the previous layer's feature vector and the cross-layer connection edge signal, and whose output is processed by the ReLU activation function. In the intra-layer aggregation operation, the weighting coefficients are calculated using the following formula:

[0074] ;

[0075] in: Let represent the weighting coefficient of the i-th topological node. This represents the centrality of the i-th topological node in the geographical layout relationship. This represents the degree of matching between the output state vector of the i-th topology node and the current input signal. This indicates the total number of topological nodes in the current layer.

[0076] It is understandable that, in terms of data comparison, after the initial coordinated seepage command sequence is generated, it can be compared with the control commands in historical rainfall events to evaluate the rationality of the multi-layer correlation network extrapolation. For example, comparing the differences in valve opening distribution between the initial coordinated seepage command sequence generated this time and the previous experience control commands under the same baseline rainfall intensity, the difference value is used to analyze the adaptability of the initial network parameter settings. The signal transmission of cross-layer connection edges adopts vector addition operation to superimpose the correlation signal into the input of the current layer; the calculation of the seepage state transition function includes linear transformation and activation function to introduce nonlinear processing capability; the dynamic adjustment of the weighting coefficients in the intra-layer aggregation rule is completed based on the real-time calculation of the matching degree between the topological node centrality and the input signal.

[0077] In practical implementation, the activation order of the environmental disturbance factor layer is fixed from the rainfall intensity layer to the vegetation cover layer, ensuring that the extrapolation process conforms to the physical transmission logic of hydrological influences. The geographical layout of modular storage and drainage units is represented by a graph structure, where nodes represent modular storage and drainage units and edges represent hydraulic connections between units. The configuration of topological nodes in the network layer is mapped according to this graph structure, ensuring that the connection relationship of topological nodes in each network layer is consistent with the actual connection of modular storage and drainage units. After the baseline rainfall intensity is imported into the rainfall intensity layer as an input source, the excitation signal propagates layer by layer along the multi-layer interconnected network. The seepage state transition function of each layer processes the input signal in combination with the node parameter set and outputs a seepage state vector. The aggregation operation within the layer is executed synchronously, integrating the output states of all topological nodes to generate aggregated state features. Finally, the initial coordinated seepage instruction sequence is spliced ​​and formatted to generate an initial sequence of instructions, which contains the target control parameters for each modular storage and drainage unit.

[0078] In one embodiment of the present invention, see [reference] Figure 3This involves determining the activation sequence of environmental disturbance factors and activating the seepage state transition function layer by layer, as well as performing intra-layer aggregation operations to generate aggregated state features. A fixed activation sequence of environmental disturbance factors is determined, starting at the rainfall intensity layer and ending at the vegetation cover layer. A baseline rainfall intensity is input to the rainfall intensity layer as the initial input excitation signal. For the current activated layer in the sequence, the output feature vector from the previous activated layer is received as the first input, and the associated signals from other related layers through all defined cross-layer connection edges are received as the second input. The first and second inputs are concatenated to form the complete input vector of the current layer. The seepage state transition function configured for the current activated layer is called to calculate the complete input vector and the node parameter set of the current activated layer, outputting the seepage state vector of the current layer. The seepage state vector output by the current layer is passed to the next layer in the activation sequence as the first input of the next layer, until all layers in the activation sequence have been calculated. After the seepage state transition function of the current layer completes the state calculation for all topological nodes, the output state vectors of all topological nodes in the current layer are obtained. According to the predefined intra-layer aggregation rules for the current layer, the weighting coefficients of the output state vectors of each topological node in the aggregation process are determined. The weighting coefficients are jointly determined by the centrality of the topological node in the geographical layout relationship and the matching degree of the current input signal. According to the weighting coefficients, the output state vectors of all topological nodes in the current layer are weighted and summed to obtain a preliminary aggregation vector. The preliminary aggregation vector is subjected to nonlinear transformation and dimension reduction to generate the aggregated state features that represent the overall seepage state features of the current layer.

[0079] In some embodiments, the example scenario is an urban drainage area containing ten modular storage and drainage units. The environmental disturbance factor activation sequence is fixed as rainfall intensity layer, surface material layer, groundwater level layer, and vegetation cover layer. The baseline rainfall intensity is 60 mm per hour, which is input to the rainfall intensity layer as the initial input excitation signal. When the surface material layer is the current active layer, it receives the output feature vector from the rainfall intensity layer as the first input and the correlation signal directly transmitted from the rainfall intensity layer through the cross-layer connection edge as the second input. The first input and the second input are concatenated to form the complete input vector of the surface material layer. The seepage state transition function of the surface material layer is called, and combined with the node parameter set of the surface material layer, the seepage state vector of the surface material layer containing ten topological nodes is calculated and output. This seepage state vector is passed to the groundwater level layer as its first input until the vegetation cover layer is calculated. For the surface material layer, after the seepage state transfer function is calculated, the output state vectors of ten topological nodes are obtained, each with a dimension of 8. According to the intra-layer aggregation rule, the weighting coefficient of the output state vector of each topological node is calculated. The weighting coefficient is determined by the degree centrality of the topological node in the modular storage and drainage unit connection graph and the cosine similarity between the output state vector of the topological node and the complete input signal of the surface material layer. The ten output state vectors are weighted and summed to obtain the preliminary aggregation vector. The preliminary aggregation vector is subjected to nonlinear transformation and dimension reduction through a fully connected layer to output the aggregated state features with a dimension of 4.

[0080] Optionally, the weighting coefficients are calculated using a normalized product form, using the following formula:

[0081] ;

[0082] in: This represents the weighting coefficients of the output state vector of the k-th topological node. The degree centrality of the k-th topological node in the geographical layout relationship is calculated by connecting the nodes in the graph using modular storage and drainage units. This represents the matching degree between the output state vector of the k-th topological node and the complete input signal of the current layer, calculated using cosine similarity. This represents the total number of topological nodes in the current layer. The activation sequence of environmental disturbance factors is preset and fixed to ensure the consistency of the network inference order. Correlated signals are transmitted through cross-layer connection edges, the definition of which is based on the physical coupling relationship between environmental disturbance factors. For example, the groundwater level layer simultaneously receives related signals from the surface material layer and the rainfall intensity layer. Nonlinear transformation and dimensionality reduction are implemented through a fully connected neural network layer with a tanh activation function, reducing the dimension of the initial aggregated vector from the dimension of the topological node output state vector to the preset dimension of the aggregated state features.

[0083] It is understandable that, in terms of data comparison, the differences in dimensionality and information representation of data before and after the intra-layer aggregation operation can be compared. For example, the output state vector of ten topological nodes in the surface material layer is a 10x8 matrix, while the aggregated state feature generated after the intra-layer aggregation operation is a 1x4 vector. Data comparison shows that the aggregated state feature achieves compression and integration of intra-layer node information. Centrality is calculated based on the topological structure of the modular storage and drainage unit connection diagram; units with more connections have higher centrality values. The matching degree is calculated in real time, depending on the current input signal and the content of the output state vector of each topological node. The weighted summation operation linearly combines all topological node output state vectors based on the calculated weighting coefficients; nonlinear transformation introduces nonlinear relationships, and dimensionality reduction reduces data dimensionality to form a compact layer feature representation. The formation of the complete input vector is achieved through vector concatenation operations, ensuring that all information from the first and second inputs is retained and sent to the seepage state transition function for processing.

[0084] In one embodiment of the present invention, dynamic mapping and response of field units are implemented, the initial coordinated seepage command sequence is parsed, and it is decoded into a control command set for each specific modular storage and discharge unit. The control command set includes target seepage rate, valve opening degree, and pumping power information. The control command set is synchronously sent to the corresponding modular storage and discharge unit at the construction site for execution through a distributed control system. During the execution of commands by the modular storage and discharge unit, physical operation data is collected in real time through sensor arrays deployed at each unit and connecting pipelines. The physical operation data includes actual flow rate, water level, and water storage volume. The physical operation data is bound with a preset timestamp and unit identifier, and after data fusion processing, structured operation status feedback information with spatial location is generated.

[0085] In some embodiments, the example scenario includes eight modular storage and drainage units. The initial coordinated seepage command sequence is an encoded data packet containing the control parameters of the eight modular storage and drainage units in sequence. The parsing process, based on a preset decoding protocol, decomposes the data packet into eight independent control command sets. Each control command set corresponds to an identifier of a modular storage and drainage unit. For example, the control command set for modular storage and drainage unit identifier A1 contains the following: target seepage rate of 0.5 cubic meters per second, valve opening of 65%, and pumping power of 2200 watts. Through the field control unit in the distributed control system, the eight control command sets are distributed to the eight modular storage and drainage units identified as A1 to A8. Upon receiving the commands, the modular storage and drainage units drive their internal regulating valves and pumps to perform actions. Simultaneously with the modular storage and drainage units executing the commands, flow rate sensors, water level sensors, and volume sensors installed inside each modular storage and drainage unit begin operation, collecting physical operating data on actual flow rate, water level, and storage volume once per second. Pressure sensors installed on the connecting pipelines simultaneously collect pipeline pressure data as auxiliary information.

[0086] Understandably, in terms of data comparison, the differences in data format before and after parsing can be demonstrated. The initial coordinated seepage command sequence, as a whole encoded package, is parsed into eight independent control command sets with clearly defined execution parameters. The raw physical operation data collected by sensors is a discrete, sensor-separated reading stream. After data fusion and binding processing, it forms structured operational status feedback information organized by modular storage and discharge unit identifiers and time series. The distributed control system ensures the time synchronization of the control command set issuance, enabling all modular storage and discharge units to begin executing commands within a coordinated time window. The deployment of the sensor array covers all key monitoring points, ensuring the comprehensiveness of physical operation data acquisition. The binding of timestamps provides a strict time sequence for subsequent analysis, and the binding of unit identifiers ensures clear spatial ownership of each data point. The setting of fusion weights is based on a priori assessment of the reliability of different sensors.

[0087] In one embodiment of the present invention, a cyclical correction of deduction and feedback is performed. The actual state of each modular storage and drainage unit in the operational status feedback information is compared with the expected target state under the benchmark rainfall intensity, and the state difference of each unit in each dimension is calculated. Based on the topological connection relationship between units, the state difference of each unit is propagated along the connection path in the multi-layer interconnected network, and the responsibility error to be borne by each topological node is calculated. According to the distribution of responsibility error, a state deviation topology diagram is constructed, which records the intensity and direction of error in each layer and node of the network. A gradient-based parameter adjustment strategy is adopted. According to the state deviation topology diagram, the node parameter set of relevant network layers in the multi-layer interconnected network is iteratively fine-tuned from top to bottom. The direction of fine-tuning is to minimize the responsibility error. After each fine-tuning, the seepage state transition function is re-run to generate a new cooperative seepage command sequence as the corrected cooperative seepage command sequence. The actual status data corresponding to each modular storage and drainage unit identifier is extracted from the operational status feedback information. The actual status data includes the actual flow velocity value, the actual water level height value, and the actual water storage volume value. Based on the benchmark rainfall intensity, the target status data corresponding to each modular storage and drainage unit identifier is retrieved from the predefined target status mapping table. The target status data includes the target flow velocity value, the target water level height value, and the target water storage volume value. For each modular storage and drainage unit, in the flow velocity dimension, the absolute difference between the actual flow velocity value and the target flow velocity value is calculated as the flow velocity status difference; in the water level height dimension, the absolute difference between the actual water level height value and the target water level height value is calculated as the water level status difference; in the water storage volume dimension, the absolute difference between the actual water storage volume value and the target water storage volume value is calculated as the volume status difference.

[0088] In some embodiments, the example scenario is a system containing six modular storage and discharge units, with a baseline rainfall intensity of 70 mm per hour. The operational status feedback information provides the actual status data of the six units. The corresponding target status data is obtained by querying the target status mapping table. Referring to Table 1, the state difference is calculated.

[0089] Table 1: Status Difference Table for Modular Storage and Discharge Units

[0090]

[0091] The topological connections between modular storage and pumping units define the error propagation path. For example, modular storage and pumping unit identifier B2 is connected to modular storage and pumping unit identifiers B1, B3, and B4. The state difference of each modular storage and pumping unit is mapped to the initial error signal of the corresponding topological node. The weights of the connecting edges are assigned according to the connection relationship and hydraulic connection strength. Diffusion calculation is performed in the directed error propagation graph, and the responsibility error of each topological node is obtained after iteration. Based on the responsibility error values ​​and direction information of all topological nodes, a state deviation topology graph is constructed. The state deviation topology graph is a data structure containing nodes and edges. In the graph, nodes represent topological nodes in a multi-layer interconnected network, and the node attributes include the responsibility error value. Edges represent connection relationships, and the edge attributes include the error propagation direction.

[0092] It is understandable that, in terms of data comparison, the calculation process of the state difference can be demonstrated, that is, by comparing the absolute difference between the actual state data and the target state data, a specific numerical state difference is obtained. The construction of the state deviation topology map is based on the distribution of responsibility error. The responsibility error and the initial state difference are numerically related but different because the responsibility error has spread through the network topology. The target state mapping table is a pre-set lookup table, the contents of which are associated with the ideal operating parameters of each modular storage and drainage unit under different baseline rainfall intensities. The spread of error along the connection path is weighted according to the strength of hydraulic connection, with paths with higher connection strength receiving more error. Gradient direction information is extracted from the state deviation topology map, indicating the direction in which the node parameter set should be adjusted to reduce responsibility error. Parameter adjustment is carried out in an iterative manner, generating a corrected coordinated seepage command sequence after each fine-tuning, and expecting to obtain operating state feedback information that is closer to the target state in the next dynamic mapping and response.

[0093] See Figure 4 This is a comparison diagram of error propagation in the water storage and drainage units of a sponge city, illustrating the difference between the initial error of a topological node and the responsibility error after propagation. The responsibility error of all units is greater than the initial error, indicating that the error propagates to related units through the topological connections between units, reflecting the "network synergy" of the sponge city's water storage and drainage system. This diagram visually demonstrates the mechanism of "error propagation" in the sponge city's water storage and drainage system, showing that the state deviation of a unit does not exist independently but affects the entire network through topological connections. This result provides a basis for subsequent "gradient parameter adjustment," requiring priority to be given to parameter correction for units with large responsibility errors (such as B5) to achieve system-level state convergence.

[0094] In one embodiment of the present invention, the specific steps of error propagation and parameter adjustment are as follows: the state difference of each modular storage and discharge unit is mapped to the topological node of the corresponding geographical location in the multi-layer interconnected network as the initial error signal of the topological node; according to the connection relationship between nodes defined in the multi-layer interconnected network, an error propagation directed graph is constructed, wherein the weight of the connection edge between nodes is assigned according to the strength of the hydraulic connection; using an iterative propagation algorithm, in the error propagation directed graph, the initial error signal of each node is weighted and distributed to the forward node along its output connection edge, while receiving the error distribution signal from the backward node along its input connection edge; after multiple iterations until the error distribution is stable, the sum of all error distribution signals received by each topological node is calculated, and the sum of the received error distribution signals is added to the initial error signal of the topological node itself, and the sum is taken as the responsibility error to be borne by the topological node. Extract the responsibility error value of each topological node and its gradient direction relative to each parameter in the node parameter set from the state deviation topology graph. Starting from the vegetation cover layer, perform parameter adjustments on each relevant network layer in reverse order. For the current adjustment layer, traverse each topological node within the adjustment layer. For each parameter in the node parameter set of the current topological node, calculate a parameter update step size based on the gradient direction information provided by the state deviation topology graph. The parameter update step size is proportional to the responsibility error value of the current node and inversely proportional to the average magnitude of historical parameter updates. Adjust the parameter values ​​in the opposite direction of the gradient direction according to the parameter update step size. After completing the parameter adjustments for all topological nodes in the current layer, recalculate the seepage state vector of the adjustment layer using the updated node parameter set and verify whether the overall responsibility error has decreased. If it has not decreased, reduce the parameter update step size proportionally and re-execute the adjustment steps for this layer until the overall responsibility error meets the minimization trend requirement, and then continue to adjust the parameters for the next layer of the network.

[0095] In some embodiments, the example scenario is a system containing five modular storage and discharge units, with state differences of D1 to D5 for the five modular storage and discharge units. The state differences D1 to D5 are mapped to five corresponding topological nodes in a multi-layer interconnected network to form initial error signals. The connection relationship between nodes is defined based on the actual pipeline connection between units and the direction of water flow. The hydraulic connection strength is calculated from the pipeline diameter and historical flow velocity data, and the weight of the connection edge is assigned accordingly to construct a directed graph for error propagation. In the iterative propagation algorithm, a propagation attenuation factor is set, and each topological node distributes its current error signal to the forward node in each iteration according to the weight ratio of the output connection edge. After five iterations, the error distribution reaches stability. The error distribution signals received by topological node N3 from topological node N2 and topological node N4 are 0.07 and 0.05, respectively. The initial error signal of topological node N3 itself is 0.12. Therefore, the error to be borne by topological node N3 is calculated as 0.07 + 0.05 + 0.12 = 0.24.

[0096] Understandably, in terms of data comparison, the changes in values ​​before and after error propagation can be shown. The initial error signal of the topology node itself, after propagation through the network, is superimposed with the received allocation signal to form the final responsibility error. The numerical distribution of the responsibility error is smoother than the initial error signal and reflects the topological correlation. Gradient direction information indicates the specific direction of adjustment for each parameter, whether it is an increase or a decrease; the average amplitude of historical parameter updates is used to adjust the current update step size to avoid oscillations during the adjustment process. The operation of proportionally reducing the parameter update step size is triggered when the overall responsibility error has not decreased, which is an adaptive adjustment mechanism; the requirement that the overall responsibility error meets the minimization trend means that the downward trend of the error is established, rather than requiring the error to be zero immediately. The construction of the directed graph of error propagation is based on predefined hydraulic connections, ensuring that the error allocation conforms to the logic of the physical process; the iterative propagation algorithm uses multiple rounds of calculation to distribute the error reasonably in the network.

[0097] See Figure 5 This is a normalized comparison chart of multi-dimensional state differences among sponge city storage and drainage units. Through normalization (unifying differences in different dimensions to the 0-1 range), it visually displays the deviation distribution of each storage and drainage unit in terms of flow velocity, water level, and volume. The core dimensions of deviation differ among different units, indicating that the state deviation of the sponge city storage and drainage system has "unit-specific" characteristics, requiring differentiated control strategies for the core deviation dimensions of different units. This chart eliminates dimensional differences through normalization, clearly comparing the degree of deviation of each unit in multiple dimensions, providing a basis for precise control of the sponge city storage and drainage system: parameters can be adjusted first for the core deviation dimensions of each unit to improve control efficiency and accuracy.

[0098] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0099] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A modular, integrated intelligent control construction method for water storage and drainage based on the concept of sponge cities, characterized in that: The method includes: A multi-layered correlation network of modular storage and discharge units is established, where each layer corresponds to an environmental disturbance factor; for the multi-layered correlation network, a set of benchmark rainfall intensities is introduced as the input source of the environmental disturbance factor. Perform inter-layer collaborative inference operation: Based on the benchmark rainfall intensity, activate the seepage state transition function corresponding to each environmental disturbance factor in the multi-layer interconnected network layer by layer to generate an initial collaborative seepage command sequence; Implement dynamic mapping and response of field units: send the initial coordinated seepage command sequence to the corresponding modular storage and discharge unit, and at the same time collect the operation status feedback information of the modular storage and discharge unit in the physical space; The simulation and feedback are used for iterative correction: Based on the operational status feedback information and the target status of the benchmark rainfall intensity, a state deviation topology is constructed. The state deviation topology is used to adjust the node parameters of the corresponding environmental disturbance factor layer in the multi-layer interconnected network in reverse order to update the seepage state transition function and generate a corrected coordinated seepage command sequence. The dynamic mapping and response steps and subsequent cyclic correction steps are executed iteratively until the running status feedback information meets the convergence condition under the benchmark rainfall intensity. The current network parameters and the final coordinated seepage command sequence are then locked to complete the pre-configuration before construction.

2. The modular integrated intelligent regulation and control construction method for water storage and drainage based on the concept of sponge cities as described in claim 1, characterized in that, The establishment of the multi-layered interconnected network for the modular storage and discharging unit specifically includes: Identify a set of environmental disturbance factors that affect the performance of the modular storage and discharge unit, including rainfall intensity, surface material, groundwater level, and vegetation cover. Each environmental disturbance factor in the set of environmental disturbance factors is assigned an independent network layer, and within each network layer, a topological node reflecting the hydraulic connection between the units is constructed based on the geographical layout relationship of the modular storage and discharge units. Between adjacent network layers, cross-layer connection edges are established based on the coupling relationship between environmental disturbance factors. These cross-layer connection edges are used to transmit the superimposed influence signal of different environmental disturbance factors on the seepage process. Each topology node is configured with a trainable set of node parameters, and an intra-layer aggregation rule is defined for each network layer to integrate input signals from lower-layer networks or cross-layer connection edges to form the output features of the corresponding network layer.

3. The modular integrated intelligent regulation and control construction method for water storage and drainage based on the concept of sponge cities as described in claim 2, characterized in that, The specific operations for performing inter-layer collaborative inference on the network include: The reference rainfall intensity is input into the network layer corresponding to the rainfall intensity as the initial excitation signal of the rainfall intensity layer; According to the preset order from the rainfall intensity layer to the vegetation coverage layer, the seepage state transfer function is activated layer by layer; for the current layer, the seepage state transfer function takes the feature vector output by the previous layer and the correlation signal passed through the cross-layer connection edge as input, and calculates it in combination with the node parameter set of the current layer to output the seepage state vector of the current layer. During the calculation process of each layer, an intra-layer aggregation operation is performed simultaneously. Based on the intra-layer aggregation rules, the output states of all topological nodes in the current layer are integrated to generate the aggregated state characteristics of the current network. The percolation state vector output by the last layer network is concatenated and formatted with the aggregated state features of all intermediate layers according to a preset encoding rule to generate the initial collaborative percolation instruction sequence.

4. The modular integrated intelligent regulation and control construction method for water storage and drainage based on the concept of sponge cities as described in claim 3, characterized in that, The specific steps for activating the seepage state transfer function layer by layer according to a preset order from the rainfall intensity layer to the vegetation cover layer include: A fixed activation sequence of environmental disturbance factors is determined, which starts at the rainfall intensity layer and ends at the vegetation cover layer; The reference rainfall intensity is input to the rainfall intensity layer as the initial input excitation signal; For the current activated layer in the sequence, the output feature vector from the previous activated layer in the sequence is received as the first input, and the correlation signal from other related layers through all defined cross-layer connection edges is received as the second input. The first input and the second input are concatenated as vectors to form the complete input vector of the current layer. The percolation state transition function configured in the current active layer is called to calculate the complete input vector and the node parameter set of the current active layer, and output the percolation state vector of the current layer. The percolation state vector output from the current layer is passed to the next layer in the activation sequence as the first input to the next layer, until all layers in the activation sequence have been calculated.

5. The modular integrated intelligent regulation and control construction method for water storage and drainage based on the concept of sponge cities according to claim 4, characterized in that, During the calculation process at each layer, an intra-layer aggregation operation is performed simultaneously. Based on the intra-layer aggregation rules, the output states of all topological nodes in the current layer are integrated to generate the aggregated state characteristics of the current network. Specific steps include: After the seepage state transition function of the current layer completes the state calculation for all topology nodes, the output state vector of all topology nodes in the current layer is obtained. Based on the predefined intra-layer aggregation rules for the current layer, the weighting coefficients of the output state vector of each topology node in the aggregation process are determined; wherein, the weighting coefficients are jointly determined by the centrality of the topology node in the geographic layout relationship and the matching degree of the current input signal. According to the weighting coefficients, the output state vectors of all topological nodes in the current layer are weighted and summed to obtain a preliminary aggregate vector. The initial aggregation vector is subjected to nonlinear transformation and dimensional reduction to generate the aggregated state features that represent the overall seepage state characteristics of the current layer.

6. The modular integrated intelligent regulation and control construction method for water storage and drainage based on the concept of sponge cities as described in claim 5, characterized in that, The dynamic mapping and response of the implementation site unit specifically includes: The initial coordinated seepage command sequence is parsed and decoded into a set of control commands for each specific modular storage and discharge unit. The set of control commands includes target seepage rate, valve opening degree and pumping power information. The control command set is synchronously sent to the corresponding modular storage and discharge unit at the construction site for execution through a distributed control system. During the execution of instructions by the modular storage and discharge unit, physical operation data is collected in real time through sensor arrays deployed at each unit and connecting pipes. The physical operation data includes actual flow rate, water level and storage volume. The physical operation data is bound to a preset timestamp and unit identifier, and after data fusion processing, structured operation status feedback information with spatial location is generated.

7. The modular integrated intelligent regulation and control construction method for water storage and drainage based on the concept of sponge cities as described in claim 6, characterized in that, The cyclical correction involving deduction and feedback specifically includes: By comparing the actual state of each modular storage and discharge unit in the operation status feedback information with the expected target state under the benchmark rainfall intensity, the state difference of each unit in each dimension is calculated. Based on the topological connection relationship between units, the state difference of each unit is propagated along the connection path in the multi-layer association network to calculate the responsibility error that each topological node should bear. Based on the distribution of the responsibility error, the state deviation topology diagram is constructed, which records the intensity and direction of the error at each layer and node of the network. A gradient-based parameter adjustment strategy is adopted. Based on the state deviation topology graph, the node parameter set of the relevant network layer in the multi-layer interconnected network is iteratively fine-tuned from top to bottom. The direction of fine-tuning is to minimize the responsibility error. After each fine-tuning, the seepage state transition function is run again to generate a new cooperative seepage command sequence as the corrected cooperative seepage command sequence.

8. The modular integrated intelligent regulation and control construction method for water storage and drainage based on the concept of sponge cities as described in claim 7, characterized in that, The specific steps for calculating the state difference of each unit in each dimension by comparing the actual state of each modular storage and discharge unit in the operational status feedback information with the expected target state under the benchmark rainfall intensity include: Extract the actual status data corresponding to the identifier of each modular storage and discharge unit from the operation status feedback information. The actual status data includes the actual flow velocity value, the actual water level height value, and the actual water storage volume value. Based on the baseline rainfall intensity, the target state data corresponding to each modular storage and drainage unit identifier is retrieved from the predefined target state mapping table. The target state data includes the target flow velocity value, the target water level height value, and the target water storage volume value. For each modular storage and discharge unit, in terms of flow velocity, the absolute difference between the actual flow velocity value and the target flow velocity value is calculated as the flow velocity state difference quantity. In the dimension of water level height, the absolute difference between the actual water level height value and the target water level height value is calculated as the water level state difference quantity; In terms of water storage volume, the absolute difference between the actual water storage volume and the target water storage volume is calculated as the volume state difference.

9. The modular integrated intelligent regulation and control construction method for water storage and drainage based on the concept of sponge city as described in claim 8, characterized in that, The specific steps for calculating the responsibility error to be borne by each topological node based on the topological connection relationship between units, propagating the state difference of each unit along the connection path in the multi-layered network, include: The state difference of each modular storage and drainage unit is mapped to the topological node of the corresponding geographical location in the multi-layer interconnected network, which serves as the initial error signal of the topological node. Based on the node connection relationships defined in the multi-layered network, an error propagation directed graph is constructed, wherein the weights of the node connection edges are assigned according to the strength of the hydraulic connection. Using an iterative propagation algorithm, in the directed error propagation graph, the initial error signal of each node is weighted and distributed to the forward nodes along its output connection edge, while the error distribution signal from the backward nodes is received along its input connection edge. After multiple iterations until the error allocation stabilizes, the sum of all error allocation signals received by each topology node is calculated, and the sum of the received error allocation signals is added to the initial error signal of the topology node itself. The sum is taken as the responsibility error that the topology node should bear.

10. The modular integrated intelligent regulation and control construction method for water storage and drainage based on the concept of sponge cities according to claim 9, characterized in that, The specific steps of employing a gradient-based parameter adjustment strategy, which iteratively fine-tunes the node parameter sets of relevant network layers in the multi-layer interconnected network from top to bottom according to the state deviation topology graph, with the fine-tuning direction being to minimize the responsibility error, include: Extract the responsibility error value of each topology node and its gradient direction relative to each parameter in the node parameter set from the state deviation topology graph; Starting from the vegetation cover layer, parameters are adjusted sequentially for each relevant network layer in reverse order; for the current adjustment layer, each topology node within the adjustment layer is traversed. For each parameter in the node parameter set of the current topology node, a parameter update step size is calculated based on the gradient direction information corresponding to the parameter provided by the state deviation topology graph; the parameter update step size is directly proportional to the responsibility error value of the current node and inversely proportional to the average amplitude of historical parameter updates. Update the step size according to the parameters, and adjust the value of the parameters in the opposite direction of the gradient direction; After all parameters of all topology nodes in the current layer have been adjusted, the percolation state vector of the adjusted layer is recalculated using the updated node parameter set, and the overall responsibility error is verified to have decreased. If it has not decreased, the parameter update step size is reduced proportionally and the adjustment steps of this layer are re-executed until the overall responsibility error meets the minimization trend requirement, and then the parameters of the next layer network are adjusted.

Images