A distributed irrigation resource regulation capability rapid evaluation method based on physical guidance deep learning

Abstract

This invention discloses a rapid assessment method for the regulation capacity of distributed irrigation resources based on physics-guided deep learning, belonging to the field of power system operation and control. It addresses the challenge of balancing computational efficiency and physical security in large-scale heterogeneous resource aggregation. The method first constructs a ground truth dataset of regulation capacity boundaries generated by the physical environment. Then, it builds a physics-guided aggregation neural network model containing modules such as individual feature encoders. The model is trained using a hybrid loss function with dynamic weights and physical constraints. Finally, an online assessment model is constructed to complete real-time evaluation. This invention also provides a corresponding assessment system, computer equipment, and storage media, enabling millisecond-level accurate and secure aggregation regulation capacity boundary assessment. This supports real-time collaborative control of virtual power plants, improves the regulation and operation efficiency of distribution networks, and contributes to the construction of new power systems.

Description

Technical Field

[0001] This invention belongs to the technical fields of power system operation control, smart grid and virtual power plant, and specifically relates to a rapid assessment method for the regulation capacity of distributed irrigation resources based on physical-guided deep learning. Background Technology

[0002] Rural power grids are rapidly transforming into clean, low-carbon, safe, and efficient new power systems. The penetration rate of new energy sources, such as wind power and solar power, in rural areas is constantly increasing. There is an urgent need to utilize virtual power plant technology to aggregate the massive distributed resources widely distributed in fields to provide regulation capacity support. Among these, distributed electric irrigation loads, as an important component of agricultural electrification, possess typical time-shifting characteristics and soil water storage capacity, making them a highly promising demand-side regulation resource. However, achieving precise aggregation of regulation capacity for large-scale heterogeneous irrigation resources faces severe challenges. While traditional aggregation methods based on Minkowski theory and mathematical programming offer high theoretical accuracy, However, its computational complexity increases exponentially with the number of devices, making it difficult to meet the real-time control requirements of power systems at the second or millisecond level. While existing pure data-driven deep learning methods are computationally fast, they typically treat the physical system as a black box, ignoring the strict differential-integral physical constraints between energy and power. This leads to the predicted regulation capacity boundary curves often exhibiting non-physical oscillations, violating the law of energy conservation, or exceeding the physical safety limits of the devices, posing scheduling safety risks. Therefore, there is an urgent need for a rapid evaluation method for distributed resource regulation capabilities that can inherit the rigor and security of physical models while possessing the millisecond-level inference speed of deep learning. Summary of the Invention

[0003] Based on the above shortcomings, this invention proposes a rapid evaluation method for the regulation capability of distributed irrigation resources based on physical-guided deep learning. It aims to solve the problem of balancing computational efficiency and physical security in large-scale heterogeneous resource aggregation, support the millisecond-level real-time collaborative regulation of massive distributed irrigation loads by virtual power plants, thereby improving the regulation capability and operating efficiency of the distribution network and contributing to the construction of new power systems.

[0004] The technical solution adopted in this invention is as follows: A rapid evaluation method for the regulation capacity of distributed irrigation resources based on physics-guided deep learning, comprising the following steps:

[0005] S1. Constructing the physical environment of distributed irrigation resources and generating a true dataset of regulation capacity boundaries: Collect static physical parameters of heterogeneous irrigation equipment, construct a dynamic environmental disturbance sequence that simulates real meteorological laws, construct a physical model of a single equipment that includes state transition equations and physical safety constraints; use a mathematical programming solver to solve the power boundary and energy boundary of the single equipment, calculate the baseline power sequence to maintain the current state of the equipment, and generate a true dataset of regulation capacity boundaries for neural network training through aggregation operations.

[0006] S2. Construct a physical-guided adjustment capability aggregation neural network model, including an individual feature encoder, a deep ensemble aggregation layer, a parameter mapping head, and a physical reconstruction layer. The individual feature encoder maps the static physical parameters of individual devices to dynamic environmental perturbation sequences as high-dimensional latent feature vectors. The deep ensemble aggregation layer performs unordered accumulation of the latent feature vectors of all individual devices through a permutation-invariant summing pooling operator to generate a global aggregation feature vector independent of the number of devices. The parameter mapping head maps the global aggregation feature vector to an aggregation power boundary sequence, an energy boundary slope sequence, a reference power sequence, and prototype correction coefficients. The physical reconstruction layer reconstructs the energy boundary curve using a cumulative integral operator and an adaptive prototype correction module, using discrete integrals of the energy boundary slope sequence, and adjusts the preset physical prototype curve to fit the energy feasible region.

[0007] S3. The neural network model is trained using a hybrid loss function training strategy based on physical constraints: the hybrid loss function includes tracking error loss and physical limit violation penalty loss, and the priority of the loss terms is adjusted by a dynamic weight adjustment strategy; the tracking error loss measures the average absolute error between the predicted boundary and the true boundary label, the physical limit violation penalty loss only generates gradient penalty when the predicted boundary exceeds the true physical feasible region, and the dynamic weight adjustment strategy adjusts the weight of the loss terms with each training round to achieve fitting the shape of the feasible region in the early stage of training and strengthening the physical safety constraints in the later stage of training.

[0008] S4. Construct an online regulation capacity assessment model to achieve real-time assessment of the regulation capacity of distributed irrigation resources: Collect real-time status information, rated operating parameters, and environmental evapotranspiration prediction data for future scheduling cycles of heterogeneous irrigation equipment, normalize them into tensor formats suitable for neural network input; call the trained neural network model to perform millisecond-level inference calculations to obtain aggregated power boundary sequences and energy boundary sequences; map the inference results into continuous time dimension power feasible domain envelopes and energy feasible domain envelopes, and visualize the adjustable potential range of the distributed irrigation resource cluster.

[0009] Furthermore, in step S1, the process of constructing the distributed irrigation resource physical environment model and generating ground truth data includes collecting static physical parameters, constructing a dynamic environmental disturbance sequence simulating real meteorological patterns, constructing a physical model of individual equipment, and using a solver to perform time-by-time optimization and aggregation calculations. The specific implementation is as follows:

[0010] The first step is to collect static physical parameters and generate dynamic environmental disturbance sequences: This involves setting the number of heterogeneous irrigation equipment clusters. With scheduling cycle For the first Each device collects its static physical parameters and constructs a dynamic environmental disturbance sequence. The static physical parameters include: maximum water storage capacity. Minimum water storage capacity Rated power Energy efficiency coefficient and initial soil moisture content The dynamic environmental disturbance sequence A sinusoidal superposition random noise model is used to simulate the real day-night cycle and meteorological uncertainties, as shown in formula (1):

[0011]

[0012] In the formula, For the first The device is Environmental evaporation water consumption at any given time; This represents the peak amplitude of daily evapotranspiration. Control the cycle period and phase of the day-night cycle respectively; Basic evaporation rate; Random noise that follows a Gaussian distribution is used to simulate random weather changes such as cloud cover. For noise variance; The function ensures that the water consumption due to evaporation is non-negative;

[0013] Then, the static physical parameters and the generated dynamic environmental disturbance sequence are input into the mathematical programming solver to construct a physical model of a single device containing constraints as shown in equations (2)-(4). Equation (2) is the physical state transition equation of the device, which describes the dynamic integral process of soil moisture under input power and environmental disturbance. Equation (3) is the device operating power constraint, which limits the device to operate within the power range allowed by the physical hardware. Equation (4) is the device water storage capacity constraint, which ensures that the soil moisture is always maintained within the safe range allowed for crop growth.

[0014]

[0015]

[0016]

[0017] In the formula, For the first A heterogeneous irrigation device in Soil water storage at any given time; This represents the soil water storage at the previous moment; The energy efficiency coefficient represents the physical efficiency of converting a unit of electrical power into available soil moisture. For the first The device is The input irrigation power at any given time; For the first The minimum operating power of each device is set to 0. Rated power of the equipment; This represents the minimum allowable water storage capacity corresponding to the crop wilting coefficient. This refers to the soil saturation water content.

[0018] After the calculation is completed, the solver is used to perform time-by-time optimization and aggregation. The specific calculation process is defined by formulas (5)-(7). Formula (5) is the boundary solution model for single-unit adjustment capability; formula (6) is the reference power calculation model; formula (7) is the distributed resource aggregation model. Based on Minkowski and the construction of the distributed resource aggregation model, through the analysis of... The boundary sequence and reference power sequence of each individual device are accumulated time-by-time to obtain the final aggregate regulation capability boundary true value label used for neural network training;

[0019]

[0020]

[0021]

[0022] In the formula, For any point in the scheduling cycle; This indicates that the problem is constrained by the aforementioned physical model; by iterating through and solving the four objective functions of formula (5), the results are obtained respectively. The device is Upper bound of power at time Lower power limit Energy Upper Boundary and lower energy level ; For the first The device is The reference power at any given time is used solely to offset ambient evapotranspiration. Without changing the theoretical input power required to store soil water; subscript This represents the aggregated cluster variables.

[0023] Furthermore, in step S2, a physical-guided adjustment capability aggregation neural network model is constructed, which is achieved through a computational process of individual feature mapping, intermediate parameter physical correction, deep set aggregation, and physical integral reconstruction. Formula (8) is the individual feature mapping model; Formula (9) is the intermediate parameter physical correction model. Split into and As input; Formula (10) is the deep set aggregation model; Formula (11) is the physical integral reconstruction model;

[0024]

[0025]

[0026]

[0027]

[0028] In the formula, For the first Each individual device feature vector is formed by splicing the static physical parameters and dynamic environmental disturbance sequence described in step S1. This is an encoding function containing a batch normalization layer and a multilayer perceptron, used to extract high-dimensional latent features; A fully connected layer for parameter decoupling; The first output of the network The original physical parameter vector of each device, including the original values ​​of power boundary, energy slope, difference width and prototype correction coefficient; This is the lower bound of the power directly predicted by the network; The original value of the predicted power range width is obtained by... The function is forced to be positive, thus ensuring the upper bound of power. ; The original value of the prototype scaling factor is corrected to obtain a value that is always positive. To prevent the physical prototype curve from flipping; A very small positive number set to prevent gradient vanishing; The first one after being corrected by formula (9) A vector of physical parameters for each individual unit; The feature transformation operator for the identity mapping before aggregation; For the summation pooling operator, for all clusters The parameter vectors of each device are accumulated bitwise to generate a global parameter vector representing the macroscopic characteristics of the cluster. ; For the final output The energy boundary value of the time-mapping process; Global parameter vector The energy boundary rate of change sequence was analyzed; This represents the accumulation operation in the time dimension, which forces the output energy curve and power curve to satisfy the differential constraint relationship; The average energy trajectory of the pre-calculated equipment cluster; , These are respectively composed of global parameter vectors The total scaling factor and total offset obtained from the analysis correspond to the sum of all individual correction factors.

[0029] Furthermore, in step S3, the training strategy based on the hybrid loss function of physical constraints is implemented through parameter iterative model updates, hybrid loss function decomposition model, course learning dynamic weight adjustment model, and asymmetric physical limit violation penalty model.

[0030] First, we construct a parameter iterative update model based on gradient descent, formalizing the training process as an iterative update of the training dataset. The extreme value optimization problem is shown in formulas (12)-(13):

[0031]

[0032]

[0033] In the formula, This is the final optimized set of parameters for the neural network obtained through the solution process. This indicates the search for the variable that minimizes the objective function. Mathematical operations; This represents the operation of summing and averaging the loss values ​​of all samples within a training batch. For sample indexes within a batch; This refers to the training batch size; The value of the mixed loss function after considering physical constraints for a single sample; The first The network prediction boundary vector and the corresponding ground truth label vector for each sample; These are the dynamic weight hyperparameters for the current training epoch. For the first Network parameters at the next iteration; The learning rate; This is the gradient vector of the total loss function with respect to the network parameters;

[0034] Formula (14) is the hybrid loss function decomposition model; Formula (15) is the course learning dynamic weight model; Formula (16) is the asymmetric physical limit violation penalty model.

[0035]

[0036]

[0037]

[0038] In the formula, This represents the L1 norm, used to quantify the underlying tracking error. These are dynamic weighting coefficients; This is a physical limit violation penalty item; This is the current training round; This is the preheating threshold; These are the initial and upper limits of the weights, respectively. It is a linear growth rate; These are the lower and upper bounds of the truly feasible region, respectively. Predict the lower and upper bounds for the network; Ensure that a penalty gradient is generated only when the prediction interval exceeds the true feasible region.

[0039] Furthermore, in step 4, the constructed online adjustment capability assessment model is implemented through the online input vector construction model, the adjustment capability boundary forward inference model, and the distributed resource adjustable domain representation model. The specific formulas are as follows: Formula (17) is the online input vector construction model; Formula (18) is the adjustment capability boundary forward inference model.

[0040]

[0041]

[0042] In the formula, These are the static nameplate parameters for the equipment; Real-time soil moisture status collected via the Internet of Things; Forecast sequence of meteorological evapotranspiration for future scheduling cycles; This is a vector concatenation operation; The parameter is fixed. Neural network inference function; The set of input vectors for all devices within the cluster; The original multidimensional time series matrix output by the network, containing power, energy, and reference values;

[0043] Formula (19) is a distributed resource adjustable domain representation model.

[0044]

[0045] In the formula, A set of distributed resource adjustable domain evaluations for the output of online inference from a neural network; These are the model predictions. Lower and upper limits of aggregate power at any given time; These are the model predictions. The lower and upper limits of aggregated energy at any given moment; This is the reference power output by the model; This indicates the extraction of the first element from the multidimensional time series matrix output by the neural network. Data from each feature channel.

[0046] This invention also provides a rapid evaluation system for the regulation capacity of distributed irrigation resources based on physics-guided deep learning, including a data acquisition module, a ground truth dataset generation module, a neural network model construction module, a model training module, an online evaluation module, and a visualization module, used to implement the process of the rapid evaluation method for the regulation capacity of distributed irrigation resources based on physics-guided deep learning as described above; wherein:

[0047] The data acquisition module is used to collect static physical parameters and real-time operating status information of heterogeneous irrigation equipment, as well as environmental evapotranspiration prediction data of the irrigation area, and to preprocess and standardize the collected multi-source data.

[0048] The truth dataset generation module is communicatively connected to the data acquisition module. It is used to construct a distributed irrigation resource physical environment model, generate a dynamic environmental disturbance sequence, build a physical model of a single device, solve the boundary parameters through a mathematical programming solver, and generate a truth dataset of the regulation capacity boundary through aggregation operations.

[0049] The neural network model building module is communicatively connected to the truth dataset generation module and is used to build a physical-guided regulation capability aggregation neural network model. The model includes an individual feature encoder, a deep ensemble aggregation layer, a parameter mapping head, and a physical reconstruction layer to realize the mapping from individual device features to the cluster aggregation regulation capability boundary.

[0050] The model training module is communicatively connected to the ground truth dataset generation module and the neural network model construction module, respectively. It is used to train the neural network model using a hybrid loss function training strategy based on physical constraints, and to optimize the training process through a dynamic weight adjustment strategy to obtain the optimal model after training.

[0051] The online evaluation module is communicatively connected to the data acquisition module and the model training module, respectively, and is used to input the normalized data collected in real time into the trained optimal model, perform millisecond-level inference calculations, and obtain the power boundary sequence and energy boundary sequence after the distributed irrigation resource cluster is aggregated.

[0052] The visualization module is communicatively connected to the online evaluation module and is used to map the inference calculation results into the power feasible domain envelope and energy feasible domain envelope in the continuous time dimension, visually displaying the adjustable potential range of the distributed irrigation resource cluster.

[0053] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method for rapid evaluation of distributed irrigation resource regulation capability based on physical guided deep learning as described above.

[0054] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for rapid evaluation of the distributed irrigation resource regulation capability based on physical-guided deep learning as described above.

[0055] Advantages and benefits of this invention: This invention constructs a deep ensemble neural network architecture including an integral reconstruction layer and combines it with a hybrid loss function training strategy that includes physical limit penalties to achieve millisecond-level accurate evaluation and secure aggregation of the regulation capability boundaries of large-scale heterogeneous devices. This invention can output aggregated regulation capability boundaries with high accuracy, physical consistency, and scheduling security within milliseconds, effectively supporting the real-time collaborative control of massive distributed resources by virtual power plants, thereby improving the regulation capability and operational efficiency of the distribution network and contributing to the construction of new power systems. Attached Figure Description

[0056] Figure 1 This is a flowchart of the method of the present invention;

[0057] Figure 2 A comparison chart of predicted polymerization power and energy with actual values. Detailed Implementation

[0058] Example 1

[0059] like Figure 1As shown, a rapid evaluation method for the regulation capacity of distributed irrigation resources based on physics-guided deep learning includes the following steps:

[0060] S1. Constructing the physical environment of distributed irrigation resources and generating a true dataset of regulation capacity boundaries: Collect static physical parameters of heterogeneous irrigation equipment, construct a dynamic environmental disturbance sequence that simulates real meteorological laws, construct a physical model of a single equipment that includes state transition equations and physical safety constraints; use a mathematical programming solver to solve the power boundary and energy boundary of the single equipment, calculate the baseline power sequence to maintain the current state of the equipment, and generate a true dataset of regulation capacity boundaries for neural network training through aggregation operations.

[0061] S2. Construct a physical-guided adjustment capability aggregation neural network model to realize the mapping from individual features to aggregation boundaries. This mainly includes an individual feature encoder for extracting high-dimensional features of the device, a deep aggregation layer for processing variable-length device inputs, a parameter mapping head for output boundary physical parameters, and a physical reconstruction layer for ensuring differential-integral physical constraints. The individual feature encoder, used to extract high-dimensional features of devices, mainly includes an input data normalization layer, a multilayer perceptron network structure, and a nonlinear activation function, used to map the static physical parameters of individual devices and the dynamic environmental perturbation sequence into high-dimensional latent feature vectors; the deep ensemble aggregation layer, used to process variable-length device inputs, mainly includes a permutation-invariant summing pooling operator, used to unorderly accumulate the latent feature vectors of all individual devices to generate a global aggregated feature vector independent of the number of devices; the parameter mapping head, used to output boundary physical parameters, mainly includes a multilayer fully connected neural network, used to map the global aggregated feature vector into an aggregated power boundary sequence, an energy boundary slope sequence, a baseline power sequence, and prototype correction coefficients; the physical reconstruction layer, used to ensure differential-integral physical constraints, mainly includes an accumulation integral operator and an adaptive prototype correction module based on statistical priors, used to reconstruct the energy boundary curve using the predicted energy boundary slope sequence through discrete integration, and to adjust the preset physical prototype curve by combining scaling coefficients and offsets to fit the final energy feasible region;

[0062] S3. A hybrid loss function training strategy based on physical constraints is proposed to guide the neural network to learn a safe and accurate feasible region. This strategy mainly includes a tracking error loss calculation module for measuring the accuracy of the prediction boundary, a physical limit violation penalty loss calculation module for constraining the safety of the prediction boundary, and a dynamic weight adjustment strategy for adjusting the priority of loss terms. Specifically, the tracking error loss calculation module for measuring the accuracy of the prediction boundary mainly includes a computational unit that calculates the average absolute error between the predicted aggregate power boundary and aggregate energy boundary and the true boundary label, guiding the model to quickly fit the overall shape and trend of the feasible region. The physical limit violation penalty loss calculation module for constraining the safety of the prediction boundary mainly includes an asymmetric error calculation unit based on a linear rectified function, which generates a gradient penalty only when the predicted minimum boundary is less than the true minimum boundary or the predicted maximum boundary is greater than the true maximum boundary, forcing the model-generated prediction boundary to converge to the inside of the true physical feasible region to ensure scheduling safety. The dynamic weight adjustment strategy for adjusting the priority of loss terms mainly includes a course learning weight coefficient update mechanism that increases with the training rounds, focusing on basic shape fitting in the early stages of training and strengthening the satisfaction of physical safety constraints in the later stages.

[0063] S4. Construct an online adjustment capacity assessment model to achieve real-time quantitative assessment of large-scale distributed resources. First, through the collection and input interface of real-time status information of distributed irrigation resources, obtain the current soil water storage, rated operating parameters, and environmental evapotranspiration prediction data of each heterogeneous irrigation device in the jurisdiction, and normalize the physical parameters into a tensor format suitable for neural network input. Then, relying on the millisecond-level inference calculation module based on the trained neural network, call the trained physical guidance neural network weights, and use integral reconstruction logic to calculate the aggregated power boundary sequence values ​​and energy boundary sequence values ​​within milliseconds. Finally, through the visualization output module of the aggregated power upper and lower limit curves and energy upper and lower limit curves, map the inferred boundary values ​​into the power feasible region envelope and energy feasible region envelope in the continuous time dimension, intuitively displaying the adjustable potential range of the distributed irrigation resource cluster in the future scheduling cycle.

[0064] In step S1, the process of constructing the physical environment model of distributed irrigation resources and generating true data is shown in formulas (1)-(7), including collecting static physical parameters and generating dynamic environmental disturbance sequences, constructing physical models of individual equipment, and using solvers to perform time-by-time optimization and aggregation operations. The specific implementation is as follows:

[0065] The first step is to collect static physical parameters and generate dynamic environmental disturbance sequences: This involves setting the number of heterogeneous irrigation equipment clusters. With scheduling cycle For the first Each device collects its static physical parameters and constructs a dynamic environmental disturbance sequence. The static physical parameters include: maximum water storage capacity. Minimum water storage capacity Rated power Energy efficiency coefficient and initial soil moisture content The dynamic environmental disturbance sequence A sinusoidal superposition random noise model is used to simulate the real day-night cycle and meteorological uncertainties, as shown in formula (1):

[0066]

[0067] In the formula, For the first The device is Environmental evaporation water consumption at any given time; This represents the peak amplitude of daily evapotranspiration. Control the cycle period and phase of the day-night cycle respectively; Basic evaporation rate; Random noise that follows a Gaussian distribution is used to simulate random weather changes such as cloud cover. For noise variance; The function ensures that the water consumption due to evaporation is non-negative.

[0068] Then, the static physical parameters and the generated dynamic environmental disturbance sequence are input into the mathematical programming solver to construct a physical model of the single device containing the constraints shown in equations (2)-(4). Equation (2) is the physical state transition equation of the device, which describes the dynamic integral process of soil moisture under input power and environmental disturbance. Equation (3) is the device operating power constraint, which limits the device to operate within the power range allowed by the physical hardware. Equation (4) is the device water storage capacity constraint, which ensures that the soil moisture is always maintained within the safe range allowed for crop growth.

[0069]

[0070]

[0071]

[0072] In the formula, For the first A heterogeneous irrigation device in Soil water storage at any given time; This represents the soil water storage at the previous moment; The energy efficiency coefficient represents the physical efficiency of converting a unit of electrical power into available soil moisture. For the first The device is The input irrigation power at any given time; For the first The minimum operating power of each device is set to 0. Rated power of the equipment; This represents the minimum allowable water storage capacity corresponding to the crop wilting coefficient. This represents the saturated water content of the soil.

[0073] After the calculation is completed, the solver is used to perform time-by-time optimization and aggregation. The specific calculation process is defined by formulas (5)-(7). Formula (5) is the boundary solution model for individual regulation capacity, which describes the extreme value optimization process under the premise of satisfying physical constraints. Formula (6) is the reference power calculation model, which describes the reference power required to maintain the current water state. Formula (7) is the distributed resource aggregation model, which describes the superposition process of cluster regulation capacity based on Minkowski sum.

[0074]

[0075]

[0076]

[0077] In the formula, For any point in the scheduling cycle; This indicates that the problem is constrained by the aforementioned physical model; by iterating through and solving the four objective functions, the results are obtained respectively. The device is Upper bound of power at time Lower power limit Energy Upper Boundary and lower energy level ; For the first The device is The reference power at any given time is used solely to offset ambient evapotranspiration. Without changing the theoretical input power required to store soil water; subscript Represents the aggregated cluster variables; through the... The boundary sequences and reference power sequences of each individual device are accumulated time-by-time to obtain the final aggregate regulation capability boundary true value label used for neural network training.

[0078] In step S2, a physical-guided adjustment capability aggregation neural network model is constructed, as defined by the computational flow in formulas (8)-(11): Formula (8) is the individual feature mapping model, describing the nonlinear encoding process from individual physical input to a high-dimensional parameter space. Formula (9) is the intermediate parameter physical correction model, which... Split into and As input, the parameter legalization process based on soft constraint functions is described. Equation (10) is a deep set aggregation model, which describes the heterogeneous resource feature fusion process implemented using the permutation invariant operator. Equation (11) is a physical integral reconstruction model, which describes the process of generating the adjustment capability boundary based on the differential-integral physical relationship.

[0079]

[0080]

[0081]

[0082]

[0083] In the formula, For the first Each individual device feature vector is formed by splicing the static physical parameters and dynamic environmental disturbance sequence described in step S1. This is an encoding function containing a batch normalization layer and a multilayer perceptron, used to extract high-dimensional latent features; A fully connected layer for parameter decoupling; The first output of the network The original physical parameter vector of each device, including the original values ​​of power boundary, energy slope, difference width and prototype correction coefficient; This is the lower bound of the power directly predicted by the network; The original value of the predicted power range width is obtained by... The function is forced to be positive, thus ensuring the upper bound of power. ; The original value of the prototype scaling factor is corrected to obtain a value that is always positive. To prevent the physical prototype curve from flipping; A very small positive number set to prevent gradient vanishing; The first one after being corrected by formula (9) A vector of physical parameters for each individual unit; The feature transformation operator for the identity mapping before aggregation; For the summation pooling operator, for all clusters The parameter vectors of each device are accumulated bitwise to generate a global parameter vector representing the macroscopic characteristics of the cluster. This operation ensures the model's adaptability to changes in the number of input devices; For the final output The energy boundary value of the time-mapping process; Global parameter vector The energy boundary rate of change sequence was analyzed; This represents the accumulation operation in the time dimension, which forces the output energy curve and power curve to satisfy the differential constraint relationship; The average energy trajectory of the pre-calculated equipment cluster; and These are respectively composed of global parameter vectors The total scaling factor and total offset obtained from the analysis correspond to the sum of all individual correction factors.

[0084] In step S3, a hybrid loss function training strategy based on physical constraints is proposed, which is implemented by the optimization model defined by formulas (12)-(16):

[0085] First, we construct a parameter iterative update model based on gradient descent, formalizing the training process as an iterative update of the training dataset. The extreme value optimization problem is shown in formulas (12)-(13):

[0086]

[0087]

[0088] In the formula, This is the final optimized set of parameters for the neural network obtained through the solution process. This indicates the search for the variable that minimizes the objective function. Mathematical operations; This represents the operation of summing and averaging the loss values ​​of all samples within a training batch, where... For sample indexes within a batch; This refers to the training batch size; The value of the mixed loss function after considering physical constraints for a single sample; The first The network prediction boundary vector and the corresponding ground truth label vector for each sample; These are the dynamic weight hyperparameters for the current training epoch. For the first Network parameters at the next iteration; The learning rate; This is the gradient vector of the total loss function with respect to the network parameters, calculated using the backpropagation algorithm.

[0089] Wherein, the hybrid loss function The computational logic of the model and its components is defined by formulas (14)-(16). Formula (14) is the hybrid loss function decomposition model. Formula (15) is the course learning dynamic weight model, which describes... Evolutionary pattern during training. Equation (16) is an asymmetric physical limit violation penalty model, which describes the unilateral safety constraint mechanism based on ReLU.

[0090]

[0091]

[0092]

[0093] In the formula, The L1 norm (mean absolute error) is used to quantify the underlying tracking error. These are dynamic weighting coefficients; This is a physical limit violation penalty item; This is the current training round; This is the preheating threshold; These are the initial and upper limits of the weights, respectively. It is a linear growth rate; For the lower and upper bounds of the real feasible domain; These are the lower and upper bounds of the network prediction, respectively. Ensure that a penalty gradient is generated only when the prediction interval exceeds the true feasible region.

[0094] In step S4, the specific process of constructing the online regulation capacity assessment model and outputting the results is realized by the reasoning and representation model defined by formulas (17)-(19):

[0095] First, a fast mapping model from the real-time data space to the adjustment capability boundary space is constructed, specifically including input vector construction and forward propagation mapping, as shown in formulas (17)-(18). Formula (17) is the online input vector construction model, which describes the feature concatenation process of real-time multi-source data. Formula (18) is the adjustment capability boundary forward inference model, which describes the process based on trained parameters. The end-to-end mapping process.

[0096]

[0097]

[0098] In the formula, These are the static nameplate parameters for the equipment; Real-time soil moisture status collected via the Internet of Things; Forecast sequence of meteorological evapotranspiration for future scheduling cycles; This is a vector concatenation operation; The parameter is fixed. Neural network inference function; The set of input vectors for all devices within the cluster; This is the original multidimensional time series matrix output by the network, containing power, energy, and reference values.

[0099] Then, the physical representation of the output is defined by formula (19). Formula (19) is a distributed resource adjustable domain representation model.

[0100]

[0101] In the formula, A set of distributed resource adjustable domain evaluations for the output of online inference from a neural network; These are the model predictions. The lower and upper limits of aggregated power at any time define the power safety range for grid dispatch based on intelligent assessment; These are the model predictions. The lower and upper limits of aggregated energy at all times characterize the online available regulation potential of the irrigation system as a "virtual energy storage" unit; This is the reference power output by the model; This indicates the extraction of the first element from the multidimensional time series matrix output by the neural network. Data from each feature channel is used to decode abstract tensor data into concrete physical evaluation metrics.

[0102] Through the above steps, a rapid, accurate, and unified quantitative assessment of the regulatory capacity boundary of massive distributed irrigation resources, while meeting physical security constraints, can be achieved.

[0103] Example 2

[0104] To verify the effectiveness of the "rapid evaluation method for the adjustment capability boundary of irrigation equipment group based on physics-guided deep learning" proposed in this invention, a simulation environment containing parameters of heterogeneous irrigation equipment and evapotranspiration disturbances was constructed, and a mathematical optimization solver was used to generate aggregate boundary truth values, thereby performing offline training and online inference verification of the neural network model.

[0105] 1. Experimental environment and prediction model configuration

[0106] The computer used for the case study analysis was configured with an AMD Ryzen 9 5900HX with Radeon Graphics, 16.0GB of RAM, and an NVIDIA GeForce RTX 3060 Lap top GPU with 4GB of VRAM. The PyTorch 2.5.1 + cu121 deep learning framework was used, with CUDA version 12.1, and the GPU was used for CUDA parallel computing to accelerate training.

[0107] 2. Example Introduction

[0108] Figure 2 (a) A comparison of the aggregated power boundaries is presented. The gray area represents the true feasible region, the black curve represents the true boundary obtained by the solver, and the red dashed line represents the prediction of this method. It can be seen that the prediction basically coincides with the true value; the predicted upper limit curve of power can follow the subtle intraday changes of the true value, with a high overall fit and no obvious overshooting, thus meeting the safety requirements of the scheduling upper limit.

[0109] Figure 2 (b) A comparison of the convergent energy boundaries is presented. It can be seen that the true upper bound rises rapidly in the early stage and then plateaus, with the predicted curve exhibiting a consistent "rapid rise-stabilization" pattern. The true lower bound is relatively flat in the first half and gradually decreases in the second half, with the prediction also maintaining a continuous and smooth evolution trend. This phenomenon is consistent with the method used in this study, which employs slope prediction and energy reconstruction through cumulative integration, enhancing the physical consistency of the energy boundary and suppressing non-physical fluctuations caused by direct regression.

[0110] The truth generation stage requires repeatedly solving the optimization problem of "single device × every time × four types of boundaries", which involves a large amount of computation and is suitable for offline data construction. However, after training, the online stage only requires one forward propagation to output a 24-hour aggregated boundary sequence, which has millisecond-level inference potential, thus meeting the real-time evaluation requirements of large-scale irrigation resource aggregated boundaries.

Claims

1. A rapid evaluation method for the distributed irrigation resource regulation capacity based on physics-guided deep learning, characterized in that, Includes the following steps: S1. Constructing the physical environment of distributed irrigation resources and generating a true dataset of regulation capacity boundaries: Collect static physical parameters of heterogeneous irrigation equipment, construct a dynamic environmental disturbance sequence that simulates real meteorological laws, construct a physical model of a single equipment that includes state transition equations and physical safety constraints; use a mathematical programming solver to solve the power boundary and energy boundary of the single equipment, calculate the baseline power sequence to maintain the current state of the equipment, and generate a true dataset of regulation capacity boundaries for neural network training through aggregation operations. S2. Construct a physical-guided adjustment capability aggregation neural network model, including an individual feature encoder, a deep ensemble aggregation layer, a parameter mapping head, and a physical reconstruction layer. The individual feature encoder maps the static physical parameters of individual devices to dynamic environmental perturbation sequences as high-dimensional latent feature vectors. The deep ensemble aggregation layer performs unordered accumulation of the latent feature vectors of all individual devices through a permutation-invariant summing pooling operator to generate a global aggregation feature vector independent of the number of devices. The parameter mapping head maps the global aggregation feature vector to an aggregation power boundary sequence, an energy boundary slope sequence, a reference power sequence, and prototype correction coefficients. The physical reconstruction layer reconstructs the energy boundary curve using a cumulative integral operator and an adaptive prototype correction module, using discrete integrals of the energy boundary slope sequence, and adjusts the preset physical prototype curve to fit the energy feasible region. S3. The neural network model is trained using a hybrid loss function training strategy based on physical constraints: the hybrid loss function includes tracking error loss and physical limit violation penalty loss, and the priority of the loss terms is adjusted by a dynamic weight adjustment strategy; the tracking error loss measures the average absolute error between the predicted boundary and the true boundary label, the physical limit violation penalty loss only generates gradient penalty when the predicted boundary exceeds the true physical feasible region, and the dynamic weight adjustment strategy adjusts the weight of the loss terms with each training round to achieve fitting the shape of the feasible region in the early stage of training and strengthening the physical safety constraints in the later stage of training. S4. Construct an online regulation capacity assessment model to achieve real-time assessment of the regulation capacity of distributed irrigation resources: Collect real-time status information, rated operating parameters, and environmental evapotranspiration prediction data for future scheduling cycles of heterogeneous irrigation equipment, normalize them into tensor formats suitable for neural network input; call the trained neural network model to perform millisecond-level inference calculations to obtain aggregated power boundary sequences and energy boundary sequences; map the inference results into continuous time dimension power feasible domain envelopes and energy feasible domain envelopes, and visualize the adjustable potential range of the distributed irrigation resource cluster.

2. The method for rapid evaluation of distributed irrigation resource regulation capacity based on physics-guided deep learning according to claim 1, characterized in that: Step S1 involves constructing the distributed irrigation resource physical environment model and generating ground truth data. This includes collecting static physical parameters, constructing a dynamic environmental disturbance sequence that simulates real meteorological patterns, constructing physical models of individual equipment, and using a solver to perform time-by-time optimization and aggregation calculations. The specific implementation is as follows: The first step is to collect static physical parameters and generate dynamic environmental disturbance sequences: This involves setting the number of heterogeneous irrigation equipment clusters. With scheduling cycle For the first Each device collects its static physical parameters and constructs a dynamic environmental disturbance sequence. The static physical parameters include: maximum water storage capacity. Minimum water storage capacity Rated power Energy efficiency coefficient and initial soil moisture content The dynamic environmental disturbance sequence A sinusoidal superposition random noise model is used to simulate the real day-night cycle and meteorological uncertainties, as shown in formula (1): In the formula, For the first The device is Environmental evaporation water consumption at any given time; This represents the peak amplitude of daily evapotranspiration. Control the cycle period and phase of the day-night cycle respectively; Basic evaporation rate; Random noise that follows a Gaussian distribution is used to simulate random weather changes such as cloud cover. For noise variance; The function ensures that the water consumption due to evaporation is non-negative; Then, the static physical parameters and the generated dynamic environmental disturbance sequence are input into the mathematical programming solver to construct a physical model of a single device containing constraints as shown in equations (2)-(4). Equation (2) is the physical state transition equation of the device, which describes the dynamic integral process of soil moisture under input power and environmental disturbance. Equation (3) is the device operating power constraint, which limits the device to operate within the power range allowed by the physical hardware. Equation (4) is the device water storage capacity constraint, which ensures that the soil moisture is always maintained within the safe range allowed for crop growth. In the formula, For the first A heterogeneous irrigation device in Soil water storage at any given time; This represents the soil water storage at the previous moment; The energy efficiency coefficient represents the physical efficiency of converting a unit of electrical power into available soil moisture. For the first The device is The input irrigation power at any given time; For the first The minimum operating power of each device is set to 0. Rated power of the equipment; This represents the minimum allowable water storage capacity corresponding to the crop wilting coefficient. This refers to the soil saturation water content. After the calculation is completed, the solver is used to perform time-by-time optimization and aggregation. The specific calculation process is defined by formulas (5)-(7). Formula (5) is the boundary solution model for single-unit adjustment capability; formula (6) is the reference power calculation model; formula (7) is the distributed resource aggregation model. Based on Minkowski and the construction of the distributed resource aggregation model, through the analysis of... The boundary sequence and reference power sequence of each individual device are accumulated time-by-time to obtain the final aggregate regulation capability boundary true value label used for neural network training; In the formula, For any point in the scheduling cycle; This indicates that the problem is constrained by the aforementioned physical model; by iterating through and solving the four objective functions of formula (5), the results are obtained respectively. The device is Upper bound of power at time Lower power limit Energy Upper Boundary and lower energy level ; For the first The device is The reference power at any given time is used solely to offset ambient evapotranspiration. Without changing the theoretical input power required to store soil water; subscript This represents the aggregated cluster variables.

3. The method for rapid evaluation of distributed irrigation resource regulation capacity based on physics-guided deep learning according to claim 2, characterized in that: In step S2, a physical-guided adjustment capability aggregation neural network model is constructed, which is achieved through the calculation process of individual feature mapping, intermediate parameter physical correction, deep set aggregation, and physical integral reconstruction. Formula (8) is the individual feature mapping model. Formula (9) is the intermediate parameter physical correction model, which will Split into and As input; Formula (10) is the deep set aggregation model; Formula (11) is the physical integral reconstruction model; In the formula, For the first Each individual device feature vector is formed by splicing the static physical parameters and dynamic environmental disturbance sequence described in step S1. This is an encoding function containing a batch normalization layer and a multilayer perceptron, used to extract high-dimensional latent features; A fully connected layer for parameter decoupling; The first output of the network The original physical parameter vector of each device, including the original values ​​of power boundary, energy slope, difference width and prototype correction coefficient; This is the lower bound of the power directly predicted by the network; The original value of the predicted power range width is obtained by... The function is forced to be positive, thus ensuring the upper bound of power. ; The original value of the prototype scaling factor is corrected to obtain a value that is always positive. To prevent the physical prototype curve from flipping; A very small positive number set to prevent gradient vanishing; The first one after being corrected by formula (9) A vector of physical parameters for each individual unit; The feature transformation operator for the identity mapping before aggregation; For the summation pooling operator, for all clusters The parameter vectors of each device are accumulated bitwise to generate a global parameter vector representing the macroscopic characteristics of the cluster. ; For the final output The energy boundary value of the time-mapping process; Global parameter vector The energy boundary rate of change sequence was analyzed; This represents the accumulation operation in the time dimension, which forces the output energy curve and power curve to satisfy the differential constraint relationship; The average energy trajectory of the pre-calculated equipment cluster; , These are respectively composed of global parameter vectors The total scaling factor and total offset obtained from the analysis correspond to the sum of all individual correction factors.

4. The method for rapid evaluation of distributed irrigation resource regulation capacity based on physics-guided deep learning according to claim 3, characterized in that: In step S3, the training strategy based on the hybrid loss function of physical constraints is implemented through parameter iterative model updates, hybrid loss function decomposition model, dynamic weight adjustment model for course learning, and asymmetric physical limit violation penalty model. First, we construct a parameter iterative update model based on gradient descent, formalizing the training process as an iterative update of the training dataset. The extreme value optimization problem is shown in formulas (12)-(13): In the formula, This is the final optimized set of parameters for the neural network obtained through the solution process. This indicates the search for the variable that minimizes the objective function. Mathematical operations; This represents the operation of summing and averaging the loss values ​​of all samples within a training batch. For sample indexes within a batch; This refers to the training batch size; The value of the mixed loss function after considering physical constraints for a single sample; The first The network prediction boundary vector and the corresponding ground truth label vector for each sample; These are the dynamic weight hyperparameters for the current training epoch. For the first Network parameters at the next iteration; The learning rate; This is the gradient vector of the total loss function with respect to the network parameters; Formula (14) is the hybrid loss function decomposition model; Formula (15) is the course learning dynamic weight model; Formula (16) is the asymmetric physical limit violation penalty model. In the formula, This represents the L1 norm, used to quantify the underlying tracking error. These are dynamic weighting coefficients; This is a physical limit violation penalty item; This is the current training round; This is the preheating threshold; These are the initial and upper limits of the weights, respectively. It is a linear growth rate; These are the lower and upper bounds of the truly feasible region, respectively. Predict the lower and upper bounds for the network; Ensure that a penalty gradient is generated only when the prediction interval exceeds the true feasible region.

5. The method for rapid evaluation of distributed irrigation resource regulation capacity based on physics-guided deep learning according to claim 4, characterized in that: In step 4, the constructed online adjustment capability assessment model is achieved through the online input vector construction model, the adjustment capability boundary forward inference model, and the distributed resource adjustable domain representation model. The specific formulas are as follows: Formula (17) is the online input vector construction model; Formula (18) is the adjustment capability boundary forward inference model. In the formula, These are the static nameplate parameters for the equipment; Real-time soil moisture status collected via the Internet of Things; Forecast sequence of meteorological evapotranspiration for future scheduling cycles; This is a vector concatenation operation; The parameter is fixed. Neural network inference function; The set of input vectors for all devices within the cluster; The original multidimensional time series matrix output by the network, containing power, energy, and reference values; Formula (19) is a distributed resource adjustable domain representation model. In the formula, A set of distributed resource adjustable domain evaluations for the output of online inference from a neural network; These are the model predictions. Lower and upper limits of aggregate power at any given time; These are the model predictions. The lower and upper limits of aggregated energy at any given moment; This is the reference power output by the model; This indicates the extraction of the first element from the multidimensional time series matrix output by the neural network. Data from each feature channel.

6. A rapid evaluation system for the distributed irrigation resource regulation capacity based on physics-guided deep learning, characterized in that, The system includes a data acquisition module, a truth dataset generation module, a neural network model construction module, a model training module, an online evaluation module, and a visualization module, used to implement the workflow of a rapid evaluation method for distributed irrigation resource regulation capacity based on physical guided deep learning as described in any one of claims 1-5; wherein: The data acquisition module is used to collect static physical parameters and real-time operating status information of heterogeneous irrigation equipment, as well as environmental evapotranspiration prediction data of the irrigation area, and to preprocess and standardize the collected multi-source data. The truth dataset generation module is communicatively connected to the data acquisition module. It is used to construct a distributed irrigation resource physical environment model, generate a dynamic environmental disturbance sequence, build a physical model of a single device, solve the boundary parameters through a mathematical programming solver, and generate a truth dataset of the regulation capacity boundary through aggregation operations. The neural network model building module is communicatively connected to the truth dataset generation module and is used to build a physical-guided regulation capability aggregation neural network model. The model includes an individual feature encoder, a deep ensemble aggregation layer, a parameter mapping head, and a physical reconstruction layer to realize the mapping from individual device features to the cluster aggregation regulation capability boundary. The model training module is communicatively connected to the ground truth dataset generation module and the neural network model construction module, respectively. It is used to train the neural network model using a hybrid loss function training strategy based on physical constraints, and to optimize the training process through a dynamic weight adjustment strategy to obtain the optimal model after training. The online evaluation module is communicatively connected to the data acquisition module and the model training module, respectively, and is used to input the normalized data collected in real time into the trained optimal model, perform millisecond-level inference calculations, and obtain the power boundary sequence and energy boundary sequence after the distributed irrigation resource cluster is aggregated. The visualization module is communicatively connected to the online evaluation module and is used to map the inference calculation results into the power feasible domain envelope and energy feasible domain envelope in the continuous time dimension, visually displaying the adjustable potential range of the distributed irrigation resource cluster.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for rapid evaluation of distributed irrigation resource regulation capability based on physical-guided deep learning as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for rapid evaluation of distributed irrigation resource regulation capability based on physical-guided deep learning as described in any one of claims 1-5.

Images