Artificial intelligence based power distribution box operational energy efficiency analysis system

Abstract

The application discloses an artificial intelligence-based power distribution box operation energy efficiency analysis system, comprising: a data processing module for collecting and preprocessing operation data of the power distribution box; a state evolution module for constructing a continuous state evolution trajectory and generating a state change path; a parameter modeling module for constructing a state embedded flow and performing parameterized modeling; a residual sharing module for inputting a state evolution tensor into a multi-channel A3C architecture; a strategy aggregation module for performing independent strategy updating and constructing a cross-channel attention graph; a record control module for controlling power distribution box load on-off logic and generating energy efficiency behavior records with timing labels; and an asynchronous updating module for inputting asynchronous gradients into a shared parameter pool for synchronous updating. The application realizes continuous modeling of the power distribution box operation state and closed-loop optimization of intelligent control decisions, significantly improving the accuracy and real-time performance of energy efficiency management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent power monitoring and energy efficiency management, and particularly relates to a power distribution box operation energy efficiency analysis system based on artificial intelligence. BACKGROUND

[0002] With the continuous development of industrial automation, building intelligence and energy digitization, as an important node of power distribution and control, the operation state and energy efficiency level of the power distribution box directly affect the stability and economy of the entire power system. The traditional power distribution box operation management method mainly relies on regular manual inspection and basic sensor monitoring. Although this method has certain practicability in the initial stage, it has obvious shortcomings in dynamicity, real-time and accuracy. Especially under the conditions of multiple loads, multiple loops and complex operation, the traditional method is difficult to comprehensively and accurately perceive the operation state of the power distribution box, and it is more difficult to realize dynamic modeling and optimization decision support of the energy efficiency level.

[0003] At present, some power distribution monitoring systems have integrated data acquisition devices and can provide basic electrical parameters such as voltage, current and power. However, the processing method of these data still stays in static means such as threshold judgment and rule reasoning. This method has two main problems: on the one hand, it cannot effectively identify the energy efficiency difference under different operating environments, resulting in weak generalization ability of optimization suggestions; on the other hand, it cannot continuously model the operation state evolution process, ignoring the behavior pattern evolution characteristics contained in the time series. More importantly, the current system lacks an intelligent analysis mechanism that integrates state recognition, behavior modeling and strategy optimization, causing the energy efficiency management strategy to often lag behind the change of the operation state, affecting the timeliness and accuracy of energy-saving decisions.

[0004] In recent years, artificial intelligence technology, especially deep learning, graph neural network and reinforcement learning, has been gradually explored and applied in power system operation optimization, providing a new direction for the intelligent upgrading of power distribution management. However, existing methods are mostly limited to isolated tasks such as classification prediction and event detection, and fail to model the overall operation trajectory, construct an interpretable and decision-making state evolution model. In addition, mainstream AI models are mostly based on discrete time point data modeling, ignoring the continuous dynamic evolution characteristics of the power distribution operation state. Neural differential equations can model continuous state changes, but lack efficient embedding and strategy feedback mechanisms; while reinforcement learning has optimization ability, but it is difficult to converge in complex control tasks with multiple objectives and structures. Therefore, how to integrate state modeling capability and strategy regulation capability to construct an intelligent energy efficiency analysis architecture with context linkage, parameter sharing and dynamic decision-making capability is still a technical bottleneck that needs to be broken through.

[0005] In addition, in the power distribution efficiency modeling process, the state interaction between system channels is often ignored. In actual operation, there is an implicit coupling relationship between different load channels and energy consumption correlation. Relying on single-channel strategy training can easily cause information fragmentation. Traditional models cannot effectively fuse the strategy difference and structure commonality among multiple channels, resulting in a decrease in the accuracy of action strategy decision-making. Currently, there is a lack of an architecture design that can dynamically introduce an attention mechanism for fusion in multiple strategy paths, so as to realize the aggregation of the load on-off strategy. At the same time, the existing technology generally uses a synchronous weight update mechanism in the model training process, which is prone to learning lag or instability when facing non-stationary and asynchronous triggered energy efficiency behavior data.

[0006] Therefore, how to provide an artificial intelligence-based power distribution box operation efficiency analysis system is a problem that those skilled in the art need to solve. SUMMARY

[0007] An object of the present application is to provide an artificial intelligence-based power distribution box operation efficiency analysis system. The present application combines neural ordinary differential equation modeling method and multi-channel A3C reinforcement learning strategy architecture, and describes in detail the continuous modeling of power distribution box operation state, cross-channel action strategy aggregation and asynchronous gradient update optimization process, which has the advantages of high state recognition accuracy, strong efficiency evaluation continuity, good control response real-time performance and strong strategy adjustment adaptability.

[0008] The artificial intelligence-based power distribution box operation efficiency analysis system according to an embodiment of the present application comprises:

[0009] A data processing module for collecting and preprocessing the operation data of the power distribution box;

[0010] A state evolution module for inputting the preprocessed operation data into a neural ordinary differential equation network, constructing a continuous state evolution trajectory, and generating a state change path with time as the independent variable;

[0011] A parameter modeling module for constructing a state embedding flow and parameterizing modeling the path change rate through a learnable control term to form a state evolution tensor;

[0012] A residual sharing module for constructing a multi-channel A3C architecture with residual connection and shared parameter pool, and inputting the state evolution tensor into all sub-channels of the main policy network and the evaluation network in the architecture;

[0013] A strategy aggregation module for performing independent strategy update in each channel and inputting a difference vector between channels to construct a cross-channel attention graph for aggregating the action selection strategy in the main policy network;

[0014] a record control module configured to control the load on-off logic of the distribution box, record the control instruction and the corresponding state change, and generate an energy efficiency behavior record with a time sequence label;

[0015] an asynchronous update module configured to generate an asynchronous gradient according to the difference between the energy efficiency behavior record and the output result of the evaluation network, and input the asynchronous gradient into a shared parameter pool in all sub-channels for synchronous update.

[0016] Optionally, the modules are realized through the following method:

[0017] S1, collecting operation data of the distribution box and performing preprocessing;

[0018] S2, inputting the preprocessed operation data into the neural differential equation network, constructing a continuous state evolution trajectory, and generating a state change path with time as the independent variable;

[0019] S3, constructing a state embedding flow based on the state change path, and parameterizing modeling the path change rate through a learnable control term to form a state evolution tensor;

[0020] S4, constructing a multi-channel A3C architecture with residual connection and a shared parameter pool, and inputting the state evolution tensor into all sub-channels of the main policy network and the evaluation network of the architecture;

[0021] S5, performing independent policy update in each channel, and constructing a cross-channel attention map with the difference vector between channels as input, for aggregating the action selection policy in the main policy network;

[0022] S6, controlling the load on-off logic of the distribution box through the action selection policy, recording the control instruction and the corresponding state change, and generating an energy efficiency behavior record with a time sequence label;

[0023] S7, generating an asynchronous gradient according to the difference between the energy efficiency behavior record and the output result of the evaluation network, and inputting the asynchronous gradient into a shared parameter pool in all sub-channels for synchronous update.

[0024] Optionally, the operation data includes voltage data, current data, active power data, reactive power data, power factor data, box temperature data, and operation duration data.

[0025] Optionally, the preprocessing includes missing value filling, outlier removal, noise smoothing, time window division, and feature normalization.

[0026] Optionally, the S2 specifically includes:

[0027] S21, constructing the preprocessed distribution box operation data into a time series data set, setting the sampling sequence as X={x1, x2,..., xn}, where x i This represents the feature vector collected at time i, which includes voltage, current, active power, reactive power, power factor, enclosure temperature, and running time. n represents the total number of sampling points in the time series.

[0028] S22. Based on the time series dataset, construct state-time pairs (t... i ,x i ), where t i Represents the time variable, x i Represents time t i The corresponding characteristic state;

[0029] S23, Pair the state-time pair (t) i ,x i The input is fed into a neural network of ordinary differential equations to construct a state evolution function h(t), which satisfies the following ordinary differential form:

[0030]

[0031] Where h(t) represents the state vector at time t, x(t) represents the feature vector at time t, f(·) represents the rate of change function, and θ represents the learnable parameter used to adjust the rate of change function;

[0032] S24. Perform continuous modeling of the state evolution function h(t) within a given time interval, and sample at fixed intervals to generate a state path sequence H = {h(t1), h(t2), ..., h(t)}. m )}, where m is the number of sampling steps, h(t) j (t) represents time point t j The state vector t j Let j represent the j-th sampling time point, satisfying j∈{1,2,...,m}.

[0033] Optionally, S3 specifically includes:

[0034] S31. Based on the state change path sequence H, construct a state embedding flow E = {e1, e2, ..., e...} m}, where each embedding vector e j It is obtained through the weighted mapping function φ(·), i.e., e j =φ(h(t) j ), where m is the number of sampling steps, satisfying j∈{1,2,...,m};

[0035] S32. Define the control weight set C = {c1, c2, ..., c...} m}, where c j Represents the embedding vector ej corresponding learnable control term;

[0036] S33, based on the state embedding flow and the control weight set, a state evolution tensor is calculated, and the parameterized modeling process is as follows:

[0037]

[0038] wherein T represents the state evolution tensor, ψ(e j ) represents a nonlinear transformation function acting on the embedding vector e j , σ(·) represents a nonlinear activation function, c j represents a control term, represents an embedding vector self-inner product result, and log2(·) represents a logarithmic function.

[0039] Optionally, the S4 specifically comprises:

[0040] S41, a multi-channel A3C architecture containing a residual connection structure and a shared parameter pool is constructed, and a sub-channel set is set as {P k , V k |k = 1, 2,..., K}, wherein P k represents the kth sub-channel of the main policy network, V k represents the kth sub-channel of the evaluation network, and K represents the number of sub-channels;

[0041] S42, a residual connection structure is constructed in each sub-channel, the input is the state evolution tensor T, the intermediate mapping is R k , and the output result is O k , and the residual connection structure is represented as O k = ρ k (T) + T, wherein ρ k (·) represents a sub-channel mapping function acting on the state evolution tensor T;

[0042] S43, the output result O k is input into the shared parameter pool, each group of shareable parameters in the parameter pool is W k , and the internal network structure of each sub-channel is represented as:

[0043]

[0044] wherein Z k represents the intermediate policy feature output by the kth channel, W k represents the shared parameter matrix used by the kth sub-channel, γ(·) represents an activation compression function, η(·) represents an embedding transformation function, and represents a Hadamard element-wise multiplication operation, represents the tensor O kThe product of the matrix is ​​tanh(·), which represents the hyperbolic tangent function.

[0045] Optionally, S5 specifically includes:

[0046] S51. In each sub-channel of the main policy network, with Z k As input, perform intra-channel policy mapping to obtain the channel action vector π. k =μ k (Z k ), where π k μ represents the action vector of the k-th sub-channel. k (·) represents the strategy generation function for the k-th sub-channel;

[0047] S52. Calculate the pairwise differences of motion vectors between each channel, and construct the channel difference vector matrix Δ, where each difference vector δ ij =π i -π j i and j represent different sub-channel numbers, Δ represents a three-dimensional difference vector array of dimension K×K×d, where K is the number of sub-channels, d is the dimension of the motion vector, and π i π represents the action vector of the i-th sub-channel. j This represents the action vector of the j-th sub-channel;

[0048] S53. Construct a cross-channel attention map A, with each set of difference vectors δ ij Using the input as input, the Softmax activation function is used to obtain α. ij =Softmax(ν(δ) ij ), where ν(·) represents the difference vector mapping function, α ij Let be the attention weight of sub-channel i for sub-channel j;

[0049] S54. Apply attention map A to all channel action vectors π of the main policy network. k Generate the fused action selection vector Π, satisfying the following equation:

[0050]

[0051] Where Π represents the fused action selection vector, K represents the number of sub-channels, and λ(Z) i Z j ) represents the strategy feature interaction coefficient between sub-channel i and sub-channel j.

[0052] Optionally, S6 specifically includes:

[0053] S61. Using the action selection vector Π as the control signal input at the current moment, determine the on / off commands for each load port of the distribution box. Let the on / off command sequence be Q = {q}.k |k=1,2,...,N}, where q k This represents the command to be executed for the k-th load port. The command value is either on or off, and N represents the number of power distribution loads.

[0054] S62. Execute the on / off command sequence Q to collect real-time state change data of each load port before and after command execution. The state change data comes from the change range of the sampling parameters in the time series dataset.

[0055] S63. Construct an energy efficiency behavior record set B = {b} k |k=1,2,...,N}, each action record b k By triple (q) k ,s k ,τ k Composed of ) where q k Indicates control command, s k Indicates the status of the distribution box after the command is executed, τ k The instruction issuance timestamp indicates that the energy efficiency behavior record set B serves as time-series tagged energy efficiency behavior data, used to describe the relationship between control actions and changes in the state of the distribution box.

[0056] Optionally, S7 specifically includes:

[0057] S71, Based on the triplet (q) in the energy efficiency behavior record set B k ,s k ,τ k ), change the status of the distribution box s k Input the k-th sub-channel in the evaluation network to obtain the corresponding evaluation value set V = {v k |k=1,2,...,N}, where v k Indicates control command q k Status of the distribution box s k The evaluation value, where N represents the number of power distribution loads;

[0058] S72, Calculation control command q k Actual energy efficiency return value r k , will r k Compared with the evaluation value v k Compare and calculate the error term g. k =r k -v k Construct an asynchronous gradient set G = {g k |k=1,2,...,N}, where g k Indicates based on behavior record b k The sub-channel asynchronous gradient is used to characterize the gap between the predicted value of the action and the actual effect at that moment;

[0059] S73, the asynchronous gradient set G is taken as an update signal, an asynchronous back propagation mechanism is used to generate an intra-channel gradient flow, and each parameter group W in the shared parameter pool is sequentially injected k , so that the main strategy network and all sub-channels in the evaluation network update the corresponding weight parameters in an asynchronous manner, and a cross-channel synchronous convergence process is realized based on the parameter pool structure, so that the overall strategy optimization direction is consistent among the distributed heterogeneous sub-channels.

[0060] The beneficial effects of the present application are:

[0061] Firstly, by introducing the neural ordinary differential equation modeling method, the present application realizes dynamic description of the operating state of the distribution box in the continuous time dimension, overcomes the defect that the traditional discrete model cannot capture the state evolution process, and effectively improves the description ability of the complex load operation trend.

[0062] Secondly, based on the multi-channel A3C architecture design, a reinforcement learning network combining residual connection and shared parameter pool is constructed, the aggregation and dynamic optimization of cross-channel strategy behavior are realized, and the problems of information fragmentation and strategy isolation among channels are effectively solved.

[0063] Finally, an asynchronous gradient feedback mechanism is constructed based on the difference between the control behavior record and the evaluation network output, which not only improves the stability of parameter update, but also enhances the response ability of the system in the non-stationary data scene, thereby realizing high-precision and high-robustness distribution box energy efficiency analysis and regulation. BRIEF DESCRIPTION OF DRAWINGS

[0064] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0065] Figure 1 a module structure diagram of the distribution box operating energy efficiency analysis system based on artificial intelligence proposed by the present application;

[0066] Figure 2 a method flowchart of the distribution box operating energy efficiency analysis system based on artificial intelligence proposed by the present application;

[0067] Figure 3 an intelligent evolution modeling and asynchronous control feedback structure diagram of the distribution box operating energy efficiency analysis system based on artificial intelligence proposed by the present application. DETAILED DESCRIPTION

[0068] The present application will now be further described in detail in conjunction with the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0069] Reference Figure 1 , the power distribution box operation energy efficiency analysis system based on artificial intelligence comprises:

[0070] A data processing module is used for collecting operation data of the power distribution box and performing preprocessing;

[0071] A state evolution module is used for inputting the preprocessed operation data into a neural ordinary differential equation network, constructing a continuous state evolution trajectory, and generating a state change path with time as the independent variable;

[0072] A parameter modeling module is used for constructing a state embedding flow and parameterizing modeling of a path change rate through a learnable control item to form a state evolution tensor;

[0073] A residual sharing module is used for constructing a multi-channel A3C architecture with residual connection and shared parameter pool, and inputting the state evolution tensor into all sub-channels of the main policy network and the evaluation network in the architecture;

[0074] A policy aggregation module is used for performing independent policy updating in each channel, inputting a difference vector between channels, constructing a cross-channel attention map, and aggregating an action selection policy in the main policy network;

[0075] A record control module is used for controlling power distribution box load on-off logic, recording control instructions and corresponding state changes, and generating energy efficiency behavior records with time sequence labels;

[0076] An asynchronous updating module is used for generating asynchronous gradients according to the difference between the energy efficiency behavior records and the output results of the evaluation network, and inputting the asynchronous gradients into the shared parameter pool in all sub-channels for synchronous updating.

[0077] The artificial intelligence analysis structure integrating the neural ordinary differential equation and the multi-channel A3C architecture is constructed, high-precision dynamic modeling of the power distribution box operation state and energy efficiency response optimization control are realized, and the continuity of energy consumption monitoring and the intelligence of control decision of the system are improved.

[0078] Reference Figures 2-3 In the embodiment, the modules are realized through the following method:

[0079] S1, collecting operation data of the power distribution box and performing preprocessing;

[0080] S2, inputting the preprocessed operation data into a neural ordinary differential equation network, constructing a continuous state evolution trajectory, and generating a state change path with time as the independent variable;

[0081] S3, constructing a state embedding flow based on the state change path, parameterizing modeling of a path change rate through a learnable control item, and forming a state evolution tensor;

[0082] S4. Construct a multi-channel A3C architecture with residual connections and a shared parameter pool, and input the state evolution tensor into all sub-channels of the master policy network and the evaluation network in the architecture respectively.

[0083] S5. Perform independent policy updates within each channel and construct a cross-channel attention graph using the inter-channel difference vector as input to aggregate action selection policies in the main policy network.

[0084] S6. Control the load switching logic of the distribution box through the action selection strategy, record the control commands and corresponding state changes, and generate energy efficiency behavior records with time sequence tags.

[0085] S7. Generate an asynchronous gradient based on the difference between the energy efficiency behavior record and the output result of the evaluation network, and input the asynchronous gradient into the shared parameter pool in all sub-channels for synchronous update.

[0086] This invention establishes a complete closed-loop process from data acquisition and state modeling to asynchronous updates through seven defined steps, enhancing the system's ability to perceive complex states and its adaptive optimization capabilities during real-time operation.

[0087] In this embodiment, the operating data includes voltage data, current data, active power data, reactive power data, power factor data, enclosure temperature data, and operating time data.

[0088] This invention expands operational data into multiple key electrical parameters, enabling comprehensive collection and high-dimensional modeling of factors influencing the energy efficiency of distribution boxes, effectively improving the input quality and completeness of state evolution modeling.

[0089] In this embodiment, the preprocessing includes missing value imputation, outlier removal, noise smoothing, time window division, and feature normalization.

[0090] This invention improves the stability and flexibility of input data and enhances the robustness of the model to abnormal fluctuations and missing noise by defining a standard data preprocessing workflow.

[0091] In this embodiment, S2 specifically includes:

[0092] S21. Construct a time series dataset from the preprocessed distribution box operation data, and let the sampling sequence be X = {x1, x2, ..., x...} n}, where x i This represents the feature vector collected at time i, which includes voltage, current, active power, reactive power, power factor, enclosure temperature, and running time. n represents the total number of sampling points in the time series.

[0093] S22. Based on the time series dataset, construct state-time pairs (t... i ,x i ), where t i Represents the time variable, x i Represents time t i The corresponding characteristic state;

[0094] S23, Pair the state-time pair (t) i ,x i The input is fed into a neural network of ordinary differential equations to construct a state evolution function h(t), which satisfies the following ordinary differential form:

[0095]

[0096] Where h(t) represents the state vector at time t, x(t) represents the feature vector at time t, f(·) represents the rate of change function, and θ represents the learnable parameter used to adjust the rate of change function;

[0097] S24. Perform continuous modeling of the state evolution function h(t) within a given time interval, and sample at fixed intervals to generate a state path sequence H = {h(t1), h(t2), ..., h(t)}. m )}, where m is the number of sampling steps, h(t) j (t) represents time point t j The state vector t j Let j represent the j-th sampling time point, satisfying j∈{1,2,...,m}.

[0098] This invention introduces a continuous modeling mechanism for state changes over time by introducing a neuron-like constant differential equation. This mechanism can capture the small but critical dynamic evolution process during load operation, providing an accurate state basis for subsequent strategy generation.

[0099] In this embodiment, S3 specifically includes:

[0100] S31. Based on the state change path sequence H, construct a state embedding flow E = {e1, e2, ..., e...} m}, where each embedding vector e j It is obtained through the weighted mapping function φ(·), i.e., e j =φ(h(t) j ), where m is the number of sampling steps, satisfying j∈{1,2,...,m};

[0101] S32. Define the control weight set C = {c1, c2, ..., c...} m}, where c j Represents the embedding vector e jCorresponding learnable control term;

[0102] S33, based on the state embedding flow and the control weight set, a state evolution tensor is calculated, and the parameterized modeling process is as follows:

[0103]

[0104] Wherein, T represents the state evolution tensor, ψ(e j ) represents a nonlinear transformation function acting on the embedding vector e j , σ(·) represents a nonlinear activation function, c j represents a control term, represents the embedding vector self-inner product result, and log2(·) represents a logarithmic function.

[0105] The application forms a state evolution tensor by constructing a deep coupling modeling relationship between the state embedding flow and the control term weight, realizes parameterized representation of the power distribution state change rate, and improves the response capability of the model to nonlinear behavior patterns.

[0106] In the embodiment, the S4 specifically comprises:

[0107] S41, a multi-channel A3C architecture containing a residual connection structure and a shared parameter pool is constructed, and a sub-channel set is set as {P k ,V k |k=1,2,...,K} is set, wherein P k represents the kth sub-channel of the main policy network, V k represents the kth sub-channel of the evaluation network, and K represents the number of sub-channels;

[0108] S42, a residual connection structure is constructed in each sub-channel, the input is the state evolution tensor T, the intermediate mapping is R k , and the output result is O k , the residual connection structure is represented as O k =ρ k (T)+T, wherein ρ k (·) represents a sub-channel mapping function acting on the state evolution tensor T;

[0109] S43, the output result O k is input into the shared parameter pool, each group of shareable parameters in the parameter pool is W k , and the internal network structure of each sub-channel is represented as:

[0110]

[0111] Wherein, Z k represents the intermediate policy feature output by the kth channel, and W kdenotes the shared parameter matrix used by the k-th sub-channel, γ(·) denotes an activation compression function, η(·) denotes an embedding transformation function, and denotes a Hadamard element-wise multiplication operation, denotes the self-matrix product form of the tensor O k , and tanh(·) denotes a hyperbolic tangent function.

[0112] The present application effectively solves the structural redundancy and information island problem between the policy network and the evaluation network by introducing a multi-channel A3C architecture with a residual structure and a shared parameter pool, and realizes efficient multi-policy channel collaborative training.

[0113] In the present embodiment, the S5 specifically includes:

[0114] S51, in each sub-channel of the main policy network, performs intra-channel policy mapping with Z k as input to obtain a channel action vector π k = μ k (Z k ), where π k denotes the action vector of the k-th sub-channel, and μ k (·) denotes a policy generation function of the k-th sub-channel;

[0115] S52, calculate the pairwise difference of action vectors between channels to construct an inter-channel difference vector matrix Δ, where each difference vector δ ij = π i - π j , i and j represent different sub-channel numbers, Δ represents a three-dimensional difference vector array with dimensions K×K×d, K is the number of sub-channels, d is the dimension of the action vector, π i denotes the action vector of the i-th sub-channel, and π j denotes the action vector of the j-th sub-channel;

[0116] S53, construct a cross-channel attention map A, with each set of difference vectors δ ij as input, and use a Softmax activation function to obtain α ij = Softmax(ν(δ ij )), where ν(·) denotes a difference vector mapping function, and α ij is the attention weight of sub-channel i to sub-channel j;

[0117] S54, apply the attention map A to all channel action vectors π k of the main policy network to generate a fused action selection vector Π, which satisfies the following formula:

[0118]

[0119] wherein, represents the fused action selection vector, K represents the number of sub-channels, and represents the policy feature interaction coefficient of sub-channel i and sub-channel j. i j

[0120] The application realizes dynamic weight fusion among multiple sub-channel strategies by constructing an inter-channel difference vector and an attention map mechanism, and improves the overall synergy and strategy accuracy of action selection.

[0121] In the embodiment, the S6 specifically includes:

[0122] S61, input the action selection vector Π as the control signal at the current time, determine the on-off instruction of each load port of the distribution box, and set the on-off instruction sequence as Q={q k |k=1, 2,..., N}, wherein q k represents the execution command corresponding to the kth load port, the command value is on or off, and N represents the number of distribution loads.

[0123] S62, execute the on-off instruction sequence Q, and collect the state change data of each load port before and after the command execution in real time, wherein the state change data is derived from the change range of the sampling parameters in the time series data set;

[0124] S63, construct an energy efficiency behavior record set B={b k |k=1, 2,..., N}, wherein each behavior record b k is composed of a triple (q k , s k , τ k ), wherein q k represents the control instruction, s k represents the state of the distribution box after the instruction execution, and τ k represents the instruction issue timestamp, and the energy efficiency behavior record set B is used as the time series labeled energy efficiency behavior data to describe the relationship between the control action and the distribution box state change.

[0125] The application realizes the causal closed-loop record between the behavior and the result by establishing the on-off control and state feedback mechanism based on the action strategy driving, and provides time series support for energy efficiency behavior tracking and strategy optimization.

[0126] In the embodiment, the S7 specifically includes:

[0127] S71, based on the triple (q k , s k , τ k ) in the energy efficiency behavior record set B, the state s k of the distribution box is determined as the time series labeled energy efficiency behavior data, and the relationship between the control action and the distribution box state change is described.​The kth sub-channel in the input evaluation network is input, and a corresponding evaluation value set V={v k |k=1,2,...,N} is obtained, wherein v k represents the control instruction q k The evaluation value of the distribution box state s k , and N represents the number of distribution loads.

[0128] S72, the control instruction q k The actual energy efficiency return value r k is calculated, r k is compared with the evaluation value v k , the error term g k =r k -v k is calculated, and the asynchronous gradient set G={g k |k=1,2,...,N} is constructed, wherein g k represents the sub-channel asynchronous gradient based on the behavior record b k , which is used to represent the gap between the action value prediction and the actual effect at this moment.

[0129] S73, the asynchronous gradient set G is used as an update signal, an asynchronous back propagation mechanism is used to generate a channel gradient flow, and each parameter group W k is injected into the parameter pool in sequence, so that the main policy network and the evaluation network update the corresponding weight parameters in an asynchronous manner, and the cross-channel synchronous convergence process is realized based on the parameter pool structure, so that the overall policy optimization direction remains consistent among the distributed heterogeneous sub-channels.

[0130] The application realizes the continuous learning ability and parameter self-adjusting ability of the policy network in the non-stationary scene by analyzing the difference between the behavior record and the evaluation result, and constructing an asynchronous gradient update mechanism, thereby improving the policy stability and energy efficiency adaptability of the system during long-term operation.

[0131] Embodiment 1:

[0132] In order to verify the feasibility of the application in the implementation, the application is applied to an integrated power distribution network in a large comprehensive plant area. The load structure of the area power distribution is complex, including multiple types of industrial production load, office lighting load and cooling unit load. There are problems such as long-term operation state fluctuation, discontinuous energy efficiency monitoring, load control response lag, etc. The traditional monitoring method is difficult to capture the fine-grained energy consumption fluctuation in time, and the unreasonable load configuration often causes instantaneous power factor drop, active power waste and current impact frequently, which affects the overall energy management efficiency and system stability.

[0133] The application is deployed in the main control platform of the high-voltage and low-voltage distribution box integrated cluster in the factory area, and through the data acquisition node with real-time data sensing capability, the voltage, current, active power, reactive power, power factor, temperature rise and running time of the distribution box are continuously collected. The above data first enters the data processing module to complete the missing value completion, abnormal value elimination and normalization processing. Then, the continuous time state evolution trajectory is generated through the neural differential equation network modeling, and the state embedding flow and state evolution tensor are constructed through the parameter modeling module, and finally input into the multi-channel A3C architecture with residual structure and shared parameter pool, and the energy efficiency strategy output and evaluation feedback are completed under the cooperation of the policy network and the evaluation network.

[0134] During operation, the main policy network outputs the on-off control vector according to the different load operating states, and records the distribution box state change data in real time after the control is completed. The evaluation network simultaneously outputs the energy efficiency value estimation of the control behavior, and the deviation between the two is used as the asynchronous update gradient backflow parameter pool to realize the self-optimization of the system. The system described in the application realizes the dynamic identification, state modeling and closed-loop management of the distribution behavior in practical application, significantly improving the load control efficiency and energy efficiency level.

[0135] In the actual scene of continuous operation for 15 days, through intelligent analysis and strategy regulation of multiple distribution circuits by the application, it is observed that the average active power utilization rate is improved from 74.28% to 89.65%, the overall system reactive power backflow is reduced by 26.87%, the average current unbalance degree is reduced by 13.42%, and the maximum instantaneous power factor is improved to 0.986. In addition, the load regulation response time is reduced from an average of 3.9 seconds to 1.6 seconds, and the power fluctuation delay response is shortened by 58.97%. In terms of strategy execution stability, after training by the asynchronous update mechanism, the strategy deviation fluctuation amplitude is stabilized within ±3.5%, which is lower than the deviation range of ±7.2% under the traditional method.

[0136] To further quantify the actual application effect of the application, the typical operation data is arranged as shown in the table:

[0137] Table 1 Comparison of distribution box operation energy efficiency data (15 days before and after implementation)

[0138]

[0139]

[0140] From the above data, it can be seen that the application not only significantly improves the distribution box operation state modeling accuracy, but also effectively reduces the energy efficiency loss and control response delay, and has stable strategy self-adaptation ability, and has wide industrial popularization value.

[0141] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An artificial intelligence based electrical panel operating energy efficiency analysis system, characterized in that, The utility model relates to a kind of energy-saving behavior record generation method based on neural ordinary differential equation, including: Data processing module, for collecting the operation data of distribution box, and pre-processing; State evolution module, for inputting the pre-processed operation data into neural ordinary differential equation network, constructing continuous state evolution trajectory, and generating state change path with time as independent variable, specifically including: The preprocessed distribution box operation data is constructed as a time series data set, and a sampling sequence is , wherein represents a feature vector collected at the th time point, including voltage value, current value, active power value, reactive power value, power factor value, box temperature value, and operation duration value, represents the total number of sampling points in the time series. Based on the time series dataset, a state time pair is constructed wherein denotes a time variable, denotes a feature state corresponding to the time t; state-time pairs are input to a neural ordinary differential equation network, constructing a state evolution function that satisfies the following ordinary differential form: ； wherein, denotes a state vector at a time point denotes a feature vector at a time point denotes a state rate of change function, denotes a learnable parameter for regulating the rate of change function;​​ State evolution function Continuous modeling is performed over a given time interval and sampled at fixed intervals to generate a sequence of state paths where is the number of sampling steps, is the time point is the state vector at time point denotes the th sampling time point, satisfying ; Parameter modeling module, for constructing state embedding flow, and parameterizing modeling path change rate through learnable control item, forming state evolution tensor, specifically including: State change path sequence based , construct state embedding flow , wherein each embedding vector , through a weighted mapping function , that is , wherein is the sampling step number, satisfying ; Setting a control weight set wherein denotes a learnable control term corresponding to an embedding vector ​ Based on state embedding flow and control weight set, state evolution tensor is calculated, and the parameterization modeling process is as follows: ； wherein, denotes the state evolution tensor, denotes a nonlinear transformation function acting on the embedding vector , denotes a nonlinear activation function, denotes a control term, denotes an embedding vector self-inner product result, denotes a logarithm function; Residual sharing module, for constructing multi-channel A3C architecture with residual connection and shared parameter pool, and inputting state evolution tensor into all sub-channels of main policy network and evaluation network in the architecture; Policy aggregation module, for executing independent policy update in each channel, and constructing cross-channel attention map with difference vector between channels as input, for aggregating action selection policy in main policy network; Record control module, for controlling distribution box load on-off logic, recording control instruction and corresponding state change, and generating energy efficiency behavior record with time sequence label; Asynchronous update module, for generating asynchronous gradient according to the difference between energy efficiency behavior record and evaluation network output result, and inputting the asynchronous gradient into shared parameter pool in all sub-channels for synchronous update.

2. The artificial intelligence based electrical panel operational energy efficiency analysis system of claim 1, wherein, The modules are realized by the following methods: S1, collect the operation data of distribution box, and pre-processing; S2, input the pre-processed operation data into neural ordinary differential equation network, construct continuous state evolution trajectory, and generate state change path with time as independent variable; S3, based on state change path, construct state embedding flow, and parameterize modeling path change rate through learnable control item, form state evolution tensor; S4, construct multi-channel A3C architecture with residual connection and shared parameter pool, and input state evolution tensor into all sub-channels of main policy network and evaluation network in the architecture; S5, execute independent policy update in each channel, and construct cross-channel attention map with difference vector between channels as input, for aggregating action selection policy in main policy network; S6, control distribution box load on-off logic through action selection policy, record control instruction and corresponding state change, and generate energy efficiency behavior record with time sequence label; S7, generate asynchronous gradient according to the difference between energy efficiency behavior record and evaluation network output result, and input the asynchronous gradient into shared parameter pool in all sub-channels for synchronous update.

3. The artificial intelligence based electrical panel operational energy efficiency analysis system of claim 2, wherein, The operation data includes voltage data, current data, active power data, reactive power data, power factor data, box temperature data and running time data.

4. The artificial intelligence based electrical panel operational energy efficiency analysis system of claim 2, wherein, The pre-processing includes missing value filling, outlier removal, noise smoothing, time window division and feature normalization.

5. The artificial intelligence based electrical panel operational energy efficiency analysis system of claim 2, wherein, The S4 specifically includes: S41, construct a multi-channel A3C architecture containing a residual connection structure and a shared parameter pool, set a sub-channel set as wherein represents the i-th sub-channel of the main policy network, represents the i-th sub-channel of the evaluation network, represents the number of sub-channels;​​ S42, constructing a residual connection structure in each sub-channel, input is a state evolution tensor , intermediate mapping is , output result is , the residual connection structure is expressed as , wherein represents a sub-channel mapping function acting on the state evolution tensor . S43, output the result Input the shared parameter pool, and each group of shareable parameters in the parameter pool is The internal network structure of each sub-channel is represented as ； wherein, denotes the intermediate policy features output by the denotes the shared parameter matrix used by the denotes the activation compression function, denotes the embedding transformation function, denotes the Hadamard element-wise multiplication operation, denotes the self-matrix multiplication form of the tensor denotes the hyperbolic tangent function.​​​ 6. The artificial intelligence based electrical panel operational energy efficiency analysis system of claim 2, wherein, The S5 specifically includes: S51, in each sub-channel of the main strategy network, with As input, perform intra-channel policy mapping to obtain the channel action vector. ,in Indicates the first Action vectors of each sub-channel Indicates the first Strategy generation function for each sub-channel; S52, calculate pairwise differences of action vectors between channels, construct a channel difference vector matrix where each difference vector , and denote different sub-channel numbers, denote a three-dimensional difference vector array with dimensions, is the number of sub-channels, is the dimension of the action vector, denotes the action vector of the th sub-channel, denotes the action vector of the th sub-channel; S53, constructing a cross-channel attention map , using each set of difference vectors as input, using an activation function to obtain , where represents a difference vector mapping function, is the attention weight for a sub-channel to the sub-channel . S54, attention map all channel action vectors applied to the main policy network , generating a fused action selection vector , satisfying the following equation: ； wherein, represents the fused action selection vector, represents the number of sub-channels, represents the sub-channels and the policy feature interaction coefficients of the sub-channels .

7. The artificial intelligence based electrical panel operational energy efficiency analysis system of claim 6, wherein, The S6 specifically includes: S61、the action selection vector As the control signal input of the current moment, the on-off instruction of each load port of the distribution box is determined, and the on-off instruction sequence is , wherein represents the execution command corresponding to the first th sub-channel, and the command value is on or off, represents the number of distribution loads; S62, execute the on-off instruction sequence S62, execute the on-off instruction sequence S62, execute the on-off instruction sequence S62, execute the on-off instruction sequence S62, execute the on-off instruction sequence S62, execute the on-off instruction sequence S62, execute the on-off instruction sequence S62, execute the on-off instruction sequence S62, execute the on-off instruction sequence S62, execute the on-off instruction sequence S62, execute the on-off S63, build a set of energy efficiency behavior records Each behavior record Composed of a triple Wherein Indicates a control instruction, Indicates the status of the distribution box after the instruction is executed, Indicates the instruction issue timestamp, the set of energy efficiency behavior records As time series tagged energy efficiency behavior data, used to describe the relationship between control actions and distribution box status changes.

8. The artificial intelligence based electrical panel operational energy efficiency analysis system of claim 7, wherein, The S7 specifically includes: S71, Based on Energy Efficiency Behavior Record Set triples in Change the status of the distribution box The input evaluation network Each sub-channel yields a corresponding set of evaluation values. ,in Indicates control commands Status of the distribution box The evaluation value, Indicates the number of power distribution loads; S72, Calculation Control Command Actual energy efficiency return value ,Will Compared with the assessment value Compare and calculate the error term. Constructing an asynchronous gradient set ,in Indicates based on behavior records The sub-channel asynchronous gradient is used to characterize the gap between the predicted value of the action and the actual effect at that moment; S73, the asynchronous gradient set As an update signal, an asynchronous back propagation mechanism is adopted to generate the gradient flow within the channel, which is sequentially injected into each parameter group in the shared parameter pool So that the main strategy network and all sub-channels in the evaluation network update their corresponding weight parameters in an asynchronous manner, and realize cross-channel synchronous convergence process based on the parameter pool structure, and ensure that the overall strategy optimization direction remains consistent among distributed heterogeneous sub-channels.

Images