Distribution box operation energy efficiency analysis system based on artificial intelligence

Abstract

The invention discloses a distribution box operation energy efficiency analysis system based on artificial intelligence, and the system comprises a data processing module which is used for collecting operation data of a distribution box and carrying out the preprocessing of the operation data; the state evolution module is used for constructing a continuous state evolution trajectory and generating a state change path; the parameter modeling module is used for constructing a state embedding stream and performing parametric modeling; the residual sharing module is used for inputting the state evolution tensor into the multi-channel A3C architecture; the strategy aggregation module is used for executing independent strategy updating and constructing a cross-channel attention map; the record control module is used for controlling the on-off logic of the load of the distribution box and generating an energy efficiency behavior record with a time sequence label; and the asynchronous updating module is used for inputting the asynchronous gradient into the shared parameter pool for synchronous updating. According to the invention, the continuous modeling of the operation state of the distribution box and the closed-loop optimization of the intelligent control decision are realized, and the accuracy and real-time performance of energy efficiency management are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent power monitoring and energy efficiency management, and in particular to an artificial intelligence-based distribution box operation energy efficiency analysis system. Background Art

[0002] With the continuous development of industrial automation, intelligent buildings, and energy digitization, distribution boxes, as crucial nodes for power distribution and control, have a direct impact on the stability and economic viability of the entire power system through their operating status and energy efficiency. Traditional distribution box operation and management methods rely primarily on regular manual inspections and basic sensor monitoring. While these methods have some initial practicality, they suffer from significant deficiencies in dynamics, real-time performance, and accuracy. Especially under complex operating conditions with multiple loads, multiple circuits, and circuits, traditional methods struggle to fully and accurately perceive the operating status of distribution boxes, making it even more difficult to dynamically model energy efficiency levels and support optimal decision-making.

[0003] Although some power distribution monitoring systems currently integrate data acquisition devices and can provide basic electrical parameter information such as voltage, current, and power, the processing of this data remains static, relying on threshold judgment and rule-based reasoning. This approach suffers from two major problems: on the one hand, it cannot effectively identify energy efficiency differences under different operating environments, resulting in weak generalization of optimization recommendations; on the other hand, it cannot continuously model the evolution of operating states, ignoring the evolutionary characteristics of behavioral patterns contained in time series. More critically, the current system lacks an intelligent analysis mechanism that integrates state recognition, behavioral modeling, and strategy optimization. As a result, energy efficiency management strategies often lag behind changes in operating states, affecting the timeliness and accuracy of energy-saving decisions.

[0004] In recent years, artificial intelligence technologies, particularly deep learning, graph neural networks, and reinforcement learning, have been gradually explored and applied in power system operation optimization, providing new directions for the intelligent upgrade of distribution management. However, existing methods are mostly limited to isolated tasks such as classification prediction and event detection, failing to build interpretable and decision-making state evolution models based on holistic modeling of operation trajectories. Furthermore, mainstream AI models are mostly based on data modeling at discrete time points, ignoring the inherently continuous dynamic evolution of distribution operation states. While neural ordinary differential equations can model continuous state changes, they lack efficient embedding and policy feedback mechanisms. While reinforcement learning has optimization capabilities, it struggles to converge in complex control tasks with multiple objectives and structures. Therefore, integrating state modeling and policy control capabilities to build an intelligent energy efficiency analysis architecture with contextual linkage, parameter sharing, and dynamic decision-making capabilities remains a technical bottleneck that urgently needs to be overcome.

[0005] In addition, in the process of power distribution energy efficiency modeling, the state interaction between system channels is often overlooked. In actual operation, there are implicit coupling relationships and energy consumption correlations between different load channels. Relying solely on single-channel strategy training can easily lead to information fragmentation. Traditional models cannot effectively integrate the strategy differences and structural commonalities between multiple channels, resulting in a decrease in the accuracy of action strategy decisions. There is also a lack of an architectural design that can dynamically introduce an attention mechanism for fusion in multiple strategy paths, making it impossible to achieve aggregate judgment on load on / off strategies. At the same time, existing technologies generally adopt a synchronous weight update mechanism during model training, which can easily cause learning lag or instability when faced with non-stationary, asynchronously triggered energy efficiency behavior data.

[0006] Therefore, how to provide an artificial intelligence-based distribution box operation energy efficiency analysis system is an urgent problem that technical personnel in this field need to solve. Summary of the Invention

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

[0008] The distribution box operation energy efficiency analysis system based on artificial intelligence according to an embodiment of the present invention includes:

[0009] Data processing module, used to collect the operating data of the distribution box and perform pre-processing;

[0010] The state evolution module is used to input the pre-processed operating data into the neural ordinary differential equation network, construct a continuous state evolution trajectory, and generate a state change path with time as the independent variable;

[0011] The parameter modeling module is used to construct the state embedding flow and parameterize the path change rate through learnable control terms to form a state evolution tensor;

[0012] The residual sharing module is used to build a multi-channel A3C architecture with residual connections and shared parameter pools, and input the state evolution tensor to all sub-channels of the main policy network and evaluation network in the architecture;

[0013] The policy aggregation module is used to perform independent policy updates in each channel and construct a cross-channel attention map with the inter-channel difference vector as input to aggregate the action selection policy in the main policy network;

[0014] Recording control module, used to control the on-off logic of the distribution box load, record control instructions and corresponding state changes, and generate energy efficiency behavior records with time tags;

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

[0016] Optionally, modules can be connected using the following methods:

[0017] S1. Collect the operating data of the distribution box and perform preprocessing;

[0018] S2. Input the pre-processed operating data into the neural ordinary differential equation network to construct a continuous state evolution trajectory and generate a state change path with time as the independent variable;

[0019] S3. Construct a state embedding flow based on the state change path, and parameterize the path change rate through learnable control terms to form a state evolution tensor;

[0020] S4. Build 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 main policy network and the evaluation network in the architecture;

[0021] S5. Perform independent policy updates in each channel and use the inter-channel difference vector as input to build a cross-channel attention graph for aggregating the action selection policy in the main policy network.

[0022] S6. Control the load on / off logic of the distribution box through the action selection strategy, record the control instructions and corresponding state changes, and generate energy efficiency behavior records with time series tags;

[0023] S7. Generate an asynchronous gradient based on the difference between the energy efficiency behavior record and the evaluation network output result, and input the asynchronous gradient into the shared parameter pool in all sub-channels for synchronous update.

[0024] Optionally, the operating data includes voltage data, current data, active power data, reactive power data, power factor data, cabinet temperature data and operating time 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, construct the pre-processed distribution box operation data into a time series data set, assuming that the sampling sequence is X = {x1, x2, ..., xn}, where x i It represents the feature vector collected at the i-th moment, including voltage value, current value, active power value, reactive power value, power factor value, box temperature value and running time value. n represents the total number of sampling points in the time series.

[0028] S22, based on the time series data set, construct the state time pair (t i ,x i ), where t i represents the time variable, x i Represents time t i The corresponding feature state;

[0029] S23, the state time pair (t i ,x i ) is input into the neural ordinary differential equation network to construct the state evolution function h(t), which satisfies the following ordinary differential form:

[0030]

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

[0032] S24, perform continuous modeling on the state evolution function h(t) within a given time interval, and generate a state path sequence H = {h(t1), h(t2), ..., h(t m )}, where m is the number of sampling steps, h(t j ) is the time point t j The state vector under t j represents the jth sampling time point, satisfying j∈{1,2,...,m}.

[0033] Optionally, the S3 specifically includes:

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

[0035] S32, set the control weight set C = {c1, c2, ..., c m}, where c j Representation and embedding vector ej The corresponding learnable control items;

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

[0037]

[0038] Among them, T represents the state evolution tensor, ψ(e j ) represents the effect on the embedding vector e j The nonlinear transformation function, σ(·) represents the nonlinear activation function, c j Represents a control item, represents the result of the self-inner product of the embedded vector, and log2(·) represents the logarithmic function.

[0039] Optionally, the S4 specifically includes:

[0040] S41, construct a multi-channel A3C architecture including a residual connection structure and a shared parameter pool, and set the sub-channel set to {P k ,V k |k=1,2,...,K}, where P k represents the kth subchannel of the main policy network, V k represents the kth subchannel of the evaluation network, and K represents the number of subchannels;

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

[0042] S43, output result O k Enter the shared parameter pool, and let each group of shared parameters in the parameter pool be W k , the internal network structure of each sub-channel is expressed as:

[0043]

[0044] Among them, Z k represents the intermediate strategy feature of the k-th channel output, W k represents the shared parameter matrix used by the k-th subchannel, γ(·) represents the activation compression function, η(·) represents the embedding transformation function, ⊙ represents the Hadamard element-wise product operation, Represents the tensor O kis the self-matrix product form of , and tanh(·) represents the hyperbolic tangent function.

[0045] Optionally, the S5 specifically includes:

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

[0047] S52, calculate the difference between the action vectors of each channel, and construct the 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 a dimension of K×K×d, K is the number of sub-channels, d is the dimension of the action vector, π i represents the action vector of the i-th sub-channel, π j 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 As input, use the Softmax activation function to get α ij =Softmax(ν(δ ij )), where ν(·) represents the difference vector mapping function, α ij is the attention weight of sub-channel i to sub-channel j;

[0049] S54. Apply the 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 formula:

[0050]

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

[0052] Optionally, the S6 specifically includes:

[0053] S61, take the action selection vector Π as the control signal input at the current moment, determine the on-off instructions of each load port of the distribution box, and set the on-off instruction sequence as Q={qk |k=1,2,...,N}, where q k Indicates the execution command corresponding to the kth load port. The command value is on or off. N represents the number of distribution loads.

[0054] S62, executing the on-off instruction sequence Q, collecting 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 variation range of the sampling parameters in the time series data set;

[0055] S63, constructing energy efficiency behavior record set B = {b k |k=1,2,...,N}, each behavior record b k By triples (q k ,s k ,τ k ), where q k Indicates control instructions, s k Indicates the status of the distribution box after the instruction is executed, τ k The energy efficiency behavior record set B is used as time-series labeled energy efficiency behavior data to describe the relationship between the control action and the distribution box state change.

[0056] Optionally, the S7 specifically includes:

[0057] S71, based on the triples (q k ,s k ,τ k ), the distribution box status s k Input the kth subchannel in the evaluation network and obtain the corresponding evaluation value set V = {v k |k=1,2,...,N}, where v k Indicates control instruction q k Status of distribution box k The evaluation value of , N represents the number of distribution loads;

[0058] S72, calculation control instruction q k The actual energy efficiency return value r k , r k With the evaluation value v k Compare and calculate the error term g k =r k -v k , construct asynchronous gradient set G = {g k |k=1,2,...,N}, where g k Indicates that based on behavior record b k The sub-channel asynchronous gradient is used to represent the gap between the action value prediction and the actual effect at that moment;

[0059] S73: Use the asynchronous gradient set G as the update signal, use the asynchronous reverse transfer mechanism to generate the gradient flow in the channel, and inject it into each parameter group W in the shared parameter pool in turn. k , so that the main policy network and all sub-channels in the evaluation network update their corresponding weight parameters asynchronously, and realize cross-channel synchronous convergence process based on the parameter pool structure, ensuring that the overall policy optimization direction remains consistent among distributed heterogeneous sub-channels.

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

[0061] First, by introducing the neural ordinary differential equation modeling method, the present invention realizes the dynamic characterization of the operating status 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 ability to characterize the operating trend of complex loads.

[0062] Secondly, based on the multi-channel A3C architecture design, a reinforcement learning network combining residual connections and shared parameter pools was constructed, which achieved the aggregation and dynamic optimization of cross-channel strategic behaviors, effectively solving the problems of information fragmentation and strategy isolation between channels.

[0063] Finally, an asynchronous gradient feedback mechanism is constructed by combining the difference between the control behavior record and the evaluation network output. This not only improves the stability of parameter updates, but also enhances the system's responsiveness in non-stationary data scenarios, thereby achieving high-precision and high-robustness distribution box energy efficiency analysis and regulation. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0066] Figure 2 This is a flow chart of the method for the distribution box operation energy efficiency analysis system based on artificial intelligence proposed by the present invention;

[0067] Figure 3 This is the intelligent evolution modeling and asynchronous control feedback structure diagram of the distribution box operation energy efficiency analysis system based on artificial intelligence proposed in this invention. DETAILED DESCRIPTION

[0068] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0069] refer to Figure 1 , the distribution box operation energy efficiency analysis system based on artificial intelligence includes:

[0070] Data processing module, used to collect the operating data of the distribution box and perform pre-processing;

[0071] The state evolution module is used to input the pre-processed operating data into the neural ordinary differential equation network, construct a continuous state evolution trajectory, and generate a state change path with time as the independent variable;

[0072] The parameter modeling module is used to construct the state embedding flow and parameterize the path change rate through learnable control terms to form a state evolution tensor;

[0073] The residual sharing module is used to build a multi-channel A3C architecture with residual connections and shared parameter pools, and input the state evolution tensor to all sub-channels of the main policy network and evaluation network in the architecture;

[0074] The policy aggregation module is used to perform independent policy updates in each channel and construct a cross-channel attention map with the inter-channel difference vector as input to aggregate the action selection policy in the main policy network;

[0075] Recording control module, used to control the on-off logic of the distribution box load, record control instructions and corresponding state changes, and generate energy efficiency behavior records with time tags;

[0076] The asynchronous update module is used to generate an asynchronous gradient based on the difference between the energy efficiency behavior record and the evaluation network output result, and input the asynchronous gradient into the shared parameter pool in all sub-channels for synchronous update.

[0077] By constructing an artificial intelligence analysis structure that integrates neural ordinary differential equations and a multi-channel A3C architecture, the present invention achieves high-precision dynamic modeling of the operating status of the distribution box and energy efficiency response optimization control, thereby improving the system's energy consumption monitoring continuity and the intelligence of control decisions.

[0078] refer to Figure 2-3 In this embodiment, the modules are connected through the following methods:

[0079] S1. Collect the operating data of the distribution box and perform preprocessing;

[0080] S2. Input the pre-processed operating data into the neural ordinary differential equation network to construct a continuous state evolution trajectory and generate a state change path with time as the independent variable;

[0081] S3. Construct a state embedding flow based on the state change path, and parameterize the path change rate through learnable control terms to form a state evolution tensor;

[0082] S4. Build 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 main policy network and the evaluation network in the architecture;

[0083] S5. Perform independent policy updates in each channel and use the inter-channel difference vector as input to build a cross-channel attention graph for aggregating the action selection policy in the main policy network.

[0084] S6. Control the load on / off logic of the distribution box through the action selection strategy, record the control instructions and corresponding state changes, and generate energy efficiency behavior records with time series tags;

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

[0086] The present invention constructs a complete closed-loop process from data acquisition, state modeling to asynchronous update by clarifying seven steps, enhancing the system's perception of complex states and adaptive optimization capabilities during real-time operation.

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

[0088] By expanding the operating data into a variety of key electrical parameter indicators, the present invention realizes the comprehensive collection and high-dimensional modeling of factors affecting the energy efficiency of distribution boxes, effectively improving the input quality and integrity of state evolution modeling.

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

[0090] The present invention improves the stability and plasticity of input data and enhances the robustness of the model to abnormal fluctuations and missing noise by defining a set of standard data preprocessing processes.

[0091] In this embodiment, S2 specifically includes:

[0092] S21, construct the pre-processed distribution box operation data into a time series data set, assuming that the sampling sequence is X = {x1, x2, ..., x n}, where x i It represents the feature vector collected at the i-th moment, including voltage value, current value, active power value, reactive power value, power factor value, box temperature value and running time value. n represents the total number of sampling points in the time series.

[0093] S22, based on the time series data set, construct the state time pair (t i ,x i ), where t i represents the time variable, x i Represents time t i The corresponding feature state;

[0094] S23, the state time pair (t i ,x i ) is input into the neural ordinary differential equation network to construct the state evolution function h(t), which satisfies the following ordinary differential form:

[0095]

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

[0097] S24, perform continuous modeling on the state evolution function h(t) within a given time interval, and generate a state path sequence H = {h(t1), h(t2), ..., h(t m )}, where m is the number of sampling steps, h(t j ) is the time point t j The state vector under t j represents the jth sampling time point, satisfying j∈{1,2,...,m}.

[0098] The present invention realizes a continuous modeling mechanism for state changes over time by introducing neural ordinary differential equations, which can capture the tiny but critical dynamic evolution process during load operation and provide 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 the state embedding flow E = {e1, e2, ..., e m}, where each embedding vector e j It is obtained by weighted mapping function φ(·), that is, e j =φ(h(t j )), where m is the number of sampling steps, satisfying j∈{1,2,...,m};

[0101] S32, set the control weight set C = {c1, c2, ..., c m}, where c j Representation and embedding vector e jThe corresponding learnable control items;

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

[0103]

[0104] Among them, T represents the state evolution tensor, ψ(e j ) represents the effect on the embedding vector e j The nonlinear transformation function, σ(·) represents the nonlinear activation function, c j Represents a control item, represents the result of the self-inner product of the embedded vector, and log2(·) represents the logarithmic function.

[0105] The present invention forms a state evolution tensor by constructing a deep coupling modeling relationship between the state embedding flow and the control item weight, realizes the parameterized characterization of the distribution state change rate, and improves the model's response ability to nonlinear behavior patterns.

[0106] In this embodiment, the S4 specifically includes:

[0107] S41, construct a multi-channel A3C architecture including a residual connection structure and a shared parameter pool, and set the sub-channel set to {P k ,V k |k=1,2,...,K}, where P k represents the kth subchannel of the main policy network, V k represents the kth subchannel of the evaluation network, and K represents the number of subchannels;

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

[0109] S43, output result O k Enter the shared parameter pool, and let each group of shared parameters in the parameter pool be W k , the internal network structure of each sub-channel is expressed as:

[0110]

[0111] Among them, Z k represents the intermediate strategy feature of the k-th channel output, W krepresents the shared parameter matrix used by the k-th subchannel, γ(·) represents the activation compression function, η(·) represents the embedding transformation function, ⊙ represents the Hadamard element-wise product operation, Represents the tensor O k is the self-matrix product form of , and tanh(·) represents the hyperbolic tangent function.

[0112] The present invention effectively solves the structural redundancy and information island problems 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-strategy channel collaborative training.

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

[0114] S51. In each sub-channel of the main strategy network, Z k As input, perform channel strategy mapping and obtain the channel action vector π k =μ k (Z k ), where π k represents the action vector of the kth subchannel, μ k (·) represents the strategy generation function of the kth subchannel;

[0115] S52, calculate the difference between the action vectors of each channel, and construct the 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 a dimension of K×K×d, K is the number of sub-channels, d is the dimension of the action vector, π i represents the action vector of the i-th sub-channel, π j represents 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, use the Softmax activation function to get α ij =Softmax(ν(δ ij )), where ν(·) represents the difference vector mapping function, 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 of the main policy network k , generate the fused action selection vector Π, satisfying the following formula:

[0118]

[0119] Among them, Π represents the fused action selection vector, K represents the number of sub-channels, λ(Z i ,Z j ) represents the strategy feature interaction coefficient between sub-channel i and sub-channel j.

[0120] The present invention realizes the dynamic weight fusion between multiple sub-channel strategies by constructing the inter-channel difference vector and attention graph mechanism, thereby improving the overall coordination and strategy accuracy of action selection.

[0121] In this embodiment, S6 specifically includes:

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

[0123] S62, executing the on-off instruction sequence Q, collecting 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 variation range of the sampling parameters in the time series data set;

[0124] S63, constructing energy efficiency behavior record set B = {b k |k=1,2,...,N}, each behavior record b k By triples (q k ,s k ,τ k ), where q k Indicates control instructions, s k Indicates the status of the distribution box after the instruction is executed, τ k The energy efficiency behavior record set B is used as time-series labeled energy efficiency behavior data to describe the relationship between the control action and the distribution box state change.

[0125] The present invention establishes an on-off control and state feedback mechanism driven by action strategies, thereby realizing a causal closed-loop record between behavior and results, and providing timing support for energy efficiency behavior tracking and strategy optimization.

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

[0127] S71, based on the triples (q k ,s k ,τ k ), the distribution box status s kInput the kth subchannel in the evaluation network and obtain the corresponding evaluation value set V = {v k |k=1,2,...,N}, where v k Indicates control instruction q k Status of distribution box k The evaluation value of , N represents the number of distribution loads;

[0128] S72, calculation control instruction q k The actual energy efficiency return value r k , r k With the evaluation value v k Compare and calculate the error term g k =r k -v k , construct asynchronous gradient set G = {g k |k=1,2,...,N}, where g k Indicates that based on behavior record b k The sub-channel asynchronous gradient is used to represent the gap between the action value prediction and the actual effect at that moment;

[0129] S73: Use the asynchronous gradient set G as the update signal, use the asynchronous reverse transfer mechanism to generate the gradient flow in the channel, and inject it into each parameter group W in the shared parameter pool in turn. k , so that the main policy network and all sub-channels in the evaluation network update their corresponding weight parameters asynchronously, and realize cross-channel synchronous convergence process based on the parameter pool structure, ensuring that the overall policy optimization direction remains consistent among distributed heterogeneous sub-channels.

[0130] The present invention constructs an asynchronous gradient update mechanism by performing differential analysis between behavior records and evaluation results, thereby realizing the continuous learning ability and parameter self-adjustment ability of the policy network in non-stationary scenarios, and improving the policy stability and energy efficiency adaptability of the system in long-term operation.

[0131] Example 1:

[0132] In order to verify the feasibility of the present invention in implementation, the present invention was applied to the integrated power distribution network within a large-scale comprehensive factory area. The power distribution load structure in this area is complex, including multiple types of industrial production loads, office lighting loads and cooling unit loads. There have long been problems such as severe fluctuations in operating status, discontinuous energy efficiency monitoring, and delayed load control response. Traditional monitoring methods are difficult to capture fine-grained energy consumption fluctuations in a timely manner. Improper load configuration often causes a decrease in instantaneous power factor, active power waste and frequent current surges, affecting the overall energy management efficiency and system stability.

[0133] The present invention is deployed in the main control platform of the integrated cluster of high-voltage and low-voltage distribution boxes in the factory area. By equipping data acquisition nodes with real-time data perception capabilities, it continuously collects multi-dimensional data such as voltage, current, active power, reactive power, power factor, temperature rise and operating time of the distribution box. The above data first enters the data processing module to complete missing value filling, outlier removal and normalization processing. Subsequently, the continuous-time state evolution trajectory is generated through neural ordinary differential equation network modeling, and the state embedding flow and state evolution tensor are constructed through the parameter modeling module. Finally, it is 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 dual-channel collaboration of the strategy network and the evaluation network.

[0134] During operation, the main strategy network outputs on-off control vectors based on the load's operating state and records distribution box status changes in real time after control is complete. The evaluation network also outputs an energy efficiency value estimate for the control behavior. The deviation between the two is used as an asynchronously updated gradient reflux parameter pool to achieve system self-optimization. In practical applications, the system described in this invention achieves closed-loop management of dynamic distribution behavior identification, state modeling, and strategy execution, significantly improving load control efficiency and energy efficiency.

[0135] In an actual scenario of 15 consecutive days of operation, the present invention performed intelligent analysis and strategic regulation on multiple distribution circuits, and observed that the average active power utilization rate increased from 74.28% to 89.65%, the overall system reactive power backflow decreased by 26.87%, the average current imbalance decreased by 13.42%, and the maximum instantaneous power factor increased to 0.986. In addition, the load regulation response time decreased from an average of 3.9 seconds to 1.6 seconds, and the power fluctuation delay response was shortened by 58.97%. In terms of strategy execution stability, after training with the asynchronous update mechanism, the strategy deviation fluctuation amplitude was stabilized within ±3.5%, which is lower than the ±7.2% deviation range under the traditional method.

[0136] In order to further quantify the practical application effect of the present invention, typical operation data are sorted as shown in the table:

[0137] Table 1 Comparative data on energy efficiency of distribution boxes (statistics 15 days before and after implementation)

[0138]

[0139]

[0140] It can be seen from the above data that the present invention not only significantly improves the accuracy of modeling the operating status of the distribution box, but also effectively reduces energy efficiency loss and control response delay. At the same time, it has stable strategy adaptation capabilities and has broad industrial promotion value.

[0141] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. The distribution box operation energy efficiency analysis system based on artificial intelligence is characterized by: include: Data processing module, used to collect the operating data of the distribution box and perform pre-processing; The state evolution module is used to input the pre-processed operating data into the neural ordinary differential equation network, construct a continuous state evolution trajectory, and generate a state change path with time as the independent variable; The parameter modeling module is used to construct the state embedding flow and parameterize the path change rate through learnable control terms to form a state evolution tensor; The residual sharing module is used to build a multi-channel A3C architecture with residual connections and shared parameter pools, and input the state evolution tensor to all sub-channels of the main policy network and evaluation network in the architecture; The policy aggregation module is used to perform independent policy updates in each channel and construct a cross-channel attention map with the inter-channel difference vector as input to aggregate the action selection policy in the main policy network; Recording control module, used to control the on-off logic of the distribution box load, record control instructions and corresponding state changes, and generate energy efficiency behavior records with time tags; The asynchronous update module is used to generate an asynchronous gradient based on the difference between the energy efficiency behavior record and the evaluation network output result, and input the asynchronous gradient into the shared parameter pool in all sub-channels for synchronous update.

2. The distribution box operation energy efficiency analysis system based on artificial intelligence according to claim 1 is characterized in that: The modules are implemented as follows: S1. Collect the operating data of the distribution box and perform preprocessing; S2. Input the pre-processed operating data into the neural ordinary differential equation network to construct a continuous state evolution trajectory and generate a state change path with time as the independent variable; S3. Construct a state embedding flow based on the state change path, and parameterize the path change rate through learnable control terms to form a state evolution tensor; S4. Build 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 main policy network and the evaluation network in the architecture; S5. Perform independent policy updates in each channel and use the inter-channel difference vector as input to build a cross-channel attention graph for aggregating the action selection policy in the main policy network. S6. Control the load on / off logic of the distribution box through the action selection strategy, record the control instructions and corresponding state changes, and generate energy efficiency behavior records with time series tags; S7. Generate an asynchronous gradient based on the difference between the energy efficiency behavior record and the evaluation network output result, and input the asynchronous gradient into the shared parameter pool in all sub-channels for synchronous update.

3. The distribution box operation energy efficiency analysis system based on artificial intelligence according to claim 2 is characterized in that: The operation data includes voltage data, current data, active power data, reactive power data, power factor data, box temperature data and operation time data.

4. The distribution box operation energy efficiency analysis system based on artificial intelligence according to claim 2 is characterized in that: The preprocessing includes missing value filling, outlier removal, noise smoothing, time window division and feature normalization.

5. The distribution box operation energy efficiency analysis system based on artificial intelligence according to claim 2 is characterized in that: The S2 specifically includes: S21, construct the pre-processed distribution box operation data into a time series data set, assuming that the sampling sequence is X = {x1, x2, ..., x n }, where x i It represents the feature vector collected at the i-th moment, including voltage value, current value, active power value, reactive power value, power factor value, box temperature value and running time value. n represents the total number of sampling points in the time series. S22, based on the time series data set, construct the state time pair (t i ,x i ), where t i represents the time variable, x i Represents time t i The corresponding feature state; S23, the state time pair (t i ,x i ) is input into the neural ordinary differential equation network to construct the state evolution function h(t), which satisfies the following ordinary differential form: Where h(t) represents the state vector at time point t, x(t) represents the feature vector at time point t, f(·) represents the state change rate function, and θ represents the learnable parameter used to adjust the change rate function; S24, perform continuous modeling on the state evolution function h(t) within a given time interval, and generate a state path sequence H = {h(t1), h(t2), ..., h(t m )}, where m is the number of sampling steps, h(t j ) is the time point t j The state vector under t j represents the jth sampling time point, satisfying j∈{1,2,...,m}.

6. The distribution box operation energy efficiency analysis system based on artificial intelligence according to claim 2 is characterized in that: The S3 specifically includes: S31, based on the state change path sequence H, construct the state embedding flow E = {e1, e2, ..., e m }, where each embedding vector e j It is obtained by weighted mapping function φ(·), that is, e j =φ(h(t j )), where m is the number of sampling steps, satisfying j∈{1,2,...,m}; S32, set the control weight set C = {c1, c2, ..., c m }, where c j Representation and embedding vector e j The corresponding learnable control items; S33. Based on the state embedding flow and the control weight set, the state evolution tensor is calculated. The parameterized modeling process is as follows: Among them, T represents the state evolution tensor, ψ(e j ) represents the effect on the embedding vector e j The nonlinear transformation function, σ(·) represents the nonlinear activation function, c j Represents a control item, represents the result of the self-inner product of the embedded vector, and log2(·) represents the logarithmic function.

7. The distribution box operation energy efficiency analysis system based on artificial intelligence according to claim 2 is characterized in that: The S4 specifically includes: S41, construct a multi-channel A3C architecture including a residual connection structure and a shared parameter pool, and set the sub-channel set to {P k ,V k |k=1,2,...,K}, where P k represents the kth subchannel of the main policy network, V k represents the kth subchannel of the evaluation network, and K represents the number of subchannels; S42, construct a residual connection structure in each sub-channel, the input is the state evolution tensor T, the intermediate mapping is R k , the output is O k , the residual connection structure is represented as O k =ρ k (T)+T, where ρ k (·) represents the subchannel mapping function acting on the state evolution tensor T; S43, output result O k Enter the shared parameter pool, and let each group of shared parameters in the parameter pool be W k , the internal network structure of each sub-channel is expressed as: Among them, Z k represents the intermediate strategy feature of the k-th channel output, W k represents the shared parameter matrix used by the k-th subchannel, γ(·) represents the activation compression function, η(·) represents the embedding transformation function, ⊙ represents the Hadamard element-wise product operation, Represents the tensor O k is the self-matrix product form of , and tanh(·) represents the hyperbolic tangent function.

8. The distribution box operation energy efficiency analysis system based on artificial intelligence according to claim 2 is characterized in that: The S5 specifically includes: S51. In each sub-channel of the main strategy network, Z k As input, perform channel strategy mapping and obtain the channel action vector π k =μ k (Z k ), where π k represents the action vector of the kth subchannel, μ k (·) represents the strategy generation function of the kth subchannel; S52, calculate the difference between the action vectors of each channel, and construct the 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 a dimension of K×K×d, K is the number of sub-channels, d is the dimension of the action vector, π i represents the action vector of the i-th sub-channel, π j represents the action vector of the jth sub-channel; S53, construct a cross-channel attention map A, with each set of difference vectors δ ij As input, use the Softmax activation function to get α ij =Softmax(ν(δ ij )), where ν(·) represents the difference vector mapping function, α ij is the attention weight of sub-channel i to sub-channel j; S54. Apply the 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 formula: Among them, Π represents the fused action selection vector, K represents the number of sub-channels, λ(Z i ,Z j ) represents the strategy feature interaction coefficient between sub-channel i and sub-channel j.

9. The distribution box operation energy efficiency analysis system based on artificial intelligence according to claim 2 is characterized in that: The S6 specifically includes: S61, take the action selection vector Π as the control signal input at the current moment, determine the on-off instructions of each load port of the distribution box, and set the on-off instruction sequence as Q={q k |k=1,2,...,N}, where q k Indicates the execution command corresponding to the kth load port. The command value is on or off. N represents the number of distribution loads. S62, executing the on-off instruction sequence Q, collecting 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 variation range of the sampling parameters in the time series data set; S63, constructing energy efficiency behavior record set B = {b k |k=1,2,...,N}, each behavior record b k By triples (q k ,s k ,τ k ), where q k Indicates control instructions, s k Indicates the status of the distribution box after the instruction is executed, τ k The energy efficiency behavior record set B is used as time-series labeled energy efficiency behavior data to describe the relationship between the control action and the distribution box state change.

10. The distribution box operation energy efficiency analysis system based on artificial intelligence according to claim 2 is characterized in that: The S7 specifically includes: S71, based on the triples (q k ,s k ,τ k ), the distribution box status s k Input the kth subchannel in the evaluation network and obtain the corresponding evaluation value set V = {v k |k=1,2,...,N}, where v k Indicates control instruction q k Status of distribution box k The evaluation value of , N represents the number of distribution loads; S72, calculation control instruction q k The actual energy efficiency return value r k , r k With the evaluation value v k Compare and calculate the error term g k =r k -v k , construct asynchronous gradient set G = {g k |k=1,2,...,N}, where g k Indicates that based on behavior record b k The sub-channel asynchronous gradient is used to represent the gap between the action value prediction and the actual effect at that moment; S73: Use the asynchronous gradient set G as the update signal, use the asynchronous reverse transfer mechanism to generate the gradient flow in the channel, and inject it into each parameter group W in the shared parameter pool in turn. k , so that the main policy network and all sub-channels in the evaluation network update their corresponding weight parameters asynchronously, and realize cross-channel synchronous convergence process based on the parameter pool structure, ensuring that the overall policy optimization direction remains consistent among distributed heterogeneous sub-channels.

Images