Substation anti-electromagnetic interference energy consumption auxiliary control system and method based on artificial intelligence

Abstract

The application provides a substation anti-electromagnetic interference energy consumption auxiliary control system and method based on artificial intelligence, and relates to the technical field of substation auxiliary control. The system comprises a sensing device array, a controlled energy consumption terminal and an intelligent regulation and control platform loaded with a deep reinforcement learning engine. The platform continuously monitors the instantaneous jump gradient of the parameters. When the gradient exceeds the environmental physical evolution limit scale, it is determined that an electromagnetic pulse has occurred and the anti-collapse sub-logic is activated. In the activation stage, the platform intercepts the distorted parameters, combines the historical parameters with the inherent thermodynamic decay constant of the substation, and deduces a virtual gradual change state sequence to replace the real parameter input into the engine. At the same time, the action authority limit is implemented on the output side of the engine, and the authority to start the high-power controlled energy consumption terminal is cancelled. The application prevents the risk of secondary power grid surge induced by the misstart of high-power equipment during extreme electromagnetic interference.

Description

Technical Field

[0001] This invention relates to the field of substation auxiliary control technology, and in particular to an artificial intelligence-based substation electromagnetic interference anti-electromagnetic interference energy consumption auxiliary control system and method. Background Technology

[0002] Substations house high-voltage busbars, transformers, and disconnect switches, among other electrical facilities. To ensure a suitable operating environment for the equipment within the substation, controlled energy-consuming auxiliary terminals such as air conditioning units, electric heaters, and fresh air systems are typically installed. Environmental sensors collect parameters such as temperature, humidity, and personnel activity. With technological advancements, artificial intelligence-based control systems are increasingly being applied to substation environmental regulation. These systems collect multi-dimensional environmental data and input it into a deep network model, allowing the algorithm to autonomously output corresponding scheduling commands to adjust the environmental parameters of the substation.

[0003] When high-voltage equipment inside a substation performs switching operations or encounters overvoltage surges, it radiates alternating electromagnetic fields and transient extreme electromagnetic pulses into the surrounding space. Under the interference of strong electromagnetic fields, various environmental sensors and their communication links deployed within the substation are susceptible to high-frequency energy coupling, causing their output physical state parameter data to deviate from the normal range and exhibit abrupt jumps that do not conform to the objective laws of physical evolution. Since intelligent control models highly depend on the accuracy of input data, distorted environmental data with strong impulse noise flowing into the state space nodes of the algorithm engine can easily cause offsets in the calculation of activation values ​​for hidden feature layers.

[0004] Misled by abnormal input data, intelligent decision-making networks are prone to outputting distorted action probability distributions and erroneous control commands. Substation auxiliary control networks incorporate various high-power energy-consuming terminals, which absorb significant starting surges from the power supply bus during cold starts. If, during extreme electromagnetic pulses, the control platform is affected by sudden parameter changes and mistakenly wakes up high-power equipment, the abnormally increased load can easily induce secondary grid surges in the substation's distribution network, thereby affecting the stable operation of the local power supply system. Currently, there is an urgent need for a technical solution to ensure the consistency of intelligent control model inputs and the safety of control actions in environments with strong electromagnetic interference. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based substation anti-electromagnetic interference energy consumption auxiliary control system and method to solve the problem that when a substation encounters transient extreme electromagnetic pulses, environmental parameters are distorted and interfere with the intelligent control model, resulting in distorted action output and erroneous activation of high-power energy-consuming terminals, thereby inducing secondary power grid surges.

[0006] In a first aspect, the present invention provides an artificial intelligence-based substation anti-electromagnetic interference energy consumption auxiliary control system, including a sensor array for collecting physical state parameters of the substation, a controlled energy consumption terminal for performing environmental adjustment actions, and an intelligent control platform for establishing a communication link with the sensor array and the controlled energy consumption terminal. The intelligent control platform is equipped with a deep reinforcement learning engine. The intelligent control platform is used to input the collected physical state parameters into the deep reinforcement learning engine and to schedule the controlled energy-consuming terminal according to the action space instructions output by the deep reinforcement learning engine. The intelligent control platform is internally deployed with physical boundary anchoring anti-crash sub-logic. The intelligent control platform continuously monitors the instantaneous gradient of the physical state parameters uploaded by the sensor array; When the instantaneous jump gradient exceeds the pre-calibrated environmental physical evolution limit scale, it is determined that a transient extreme electromagnetic pulse has occurred inside the substation, and the intelligent control platform triggers the physical boundary anchoring anti-collapse sub-logic. When the anti-collapse sub-logic anchored at the physical boundary is in the active phase, the intelligent control platform intercepts distorted physical state parameters and prevents distorted data from interfering with the input layer of the deep reinforcement learning engine. At the same time, the intelligent control platform retrieves historical state parameters before the pulse occurs and, in conjunction with the inherent thermodynamic decay constant of the substation building, deduces and generates a virtual slowly changing state sequence. The virtual slowly changing state sequence is then used to replace the real parameters and input into the deep reinforcement learning engine. Furthermore, during the activation phase, the intelligent control platform restricts the action permissions of the output side of the deep reinforcement learning engine, cancels the permission of the deep reinforcement learning engine to wake up or start high-power controlled energy-consuming terminals, and only allows downward control commands to maintain the current operating state or reduce the operating load, so as to prevent secondary power grid surges caused by the accidental start of high-power equipment during extreme electromagnetic pulses.

[0007] Optionally, the sensing device array includes a temperature detection unit with anti-electromagnetic interference characteristics, a presence sensing unit including millimeter-wave radar, and a visual sensing unit loaded with artificial intelligence image recognition algorithms. The communication link is built on broadband power line carrier technology, and the sensor array transmits the physical state parameters, after edge filtering preprocessing, to the intelligent control platform through the communication link.

[0008] Optionally, the deep reinforcement learning engine is built on a deep Q-network architecture; During normal operation, the deep reinforcement learning engine receives the physical state parameters after interval standardization, the multi-sensor fusion trust index, and the real-time energy consumption value of the device as state space elements. The deep reinforcement learning engine optimizes policies through a multi-objective weighted reward function, which is composed of a fusion data accuracy reward component, an energy consumption reduction reward component, and an anti-interference action execution cost penalty component aggregated according to dynamic matching parameters.

[0009] Optionally, the controlled energy-consuming terminal includes an array of intelligent lighting fixtures; The intelligent control platform runs a target detection model based on the video stream uploaded by the visual perception unit to determine whether there is an unauthorized intrusion or fire incident inside the substation. When an unauthorized intrusion or fire is detected, the intelligent control platform outputs an emergency dimming command, triggering the intelligent lighting array to activate its maximum rated brightness illumination mode, and simultaneously pushes an alarm message to the external security control center.

[0010] Optionally, the controlled energy-consuming terminal includes a smart air conditioning unit, an electric heater, and a fresh air exchange system; The intelligent control platform has an embedded logic for comparing indoor and outdoor temperatures; when it is determined that the outdoor natural temperature meets the cooling requirements of the substation and the activity trajectory of the indoor staff is in a safe isolation zone, the intelligent control platform will give priority to issuing an instruction to start the fresh air exchange system to introduce outdoor cold air. When it is determined that introducing outdoor cold air source cannot achieve the cooling target, the intelligent control platform switches control commands to start the intelligent air conditioning unit to perform supplementary cooling tasks.

[0011] Optionally, before inputting the physical state parameters into the deep reinforcement learning engine, the intelligent control platform uses a built-in interference level assessment module to calculate the statistical variance parameter of the physical state parameters in order to perform adaptive noise reduction. When Gaussian noise is detected within the preset reference band range for the statistical variance parameter, the interference level assessment module schedules the Kalman filter algorithm to smooth the data sequence. When electromagnetic interference is detected where the statistical variance parameter exceeds the upper limit of the reference band range, the interference degree assessment module switches to and calls a composite noise reduction model assembled from sequential convolutional layers and bidirectional long short-term memory network layers to perform deep reconstruction of the damaged environmental data features.

[0012] Optionally, the specific action mechanism by which the intelligent control platform monitors the instantaneous gradient jump is as follows: The numerical difference of the physical state parameter between two adjacent sampling periods is extracted at the microsecond level, and the numerical difference is divided by the sampling period duration to obtain the current jump slope. Compare the current jump slope with the maximum possible slope of natural environmental change recorded on the environmental physical evolution limit scale. When the current transition slope reaches a multiple of the set trigger threshold of the maximum possible environmental natural change slope, and at the same time the packet loss rate of the underlying broadband power line carrier communication increases and exceeds the set communication health warning baseline, the occurrence of the transient extreme electromagnetic pulse is finally confirmed.

[0013] Optionally, the specific action mechanism for generating the virtual slowly changing state sequence through deduction is as follows: Extract the last recorded effective indoor temperature value and effective outdoor temperature value before the transient extreme electromagnetic pulse occurs, and obtain the thermal resistance constant of the outer wall of the substation building, the effective heat exchange surface area parameter of the substation building, and the indoor air specific heat capacity parameter. Substituting the thermal barrier constant, the effective heat exchange surface area parameter, the indoor air specific heat capacity parameter, and the indoor-outdoor temperature difference into the preset thermodynamic hysteresis decay formula, the natural temperature drift step size under the condition of losing environmental control measures is calculated. Based on the natural temperature drift step size, the virtual slowly changing state sequence is accumulated on the time axis.

[0014] Optionally, the intelligent control platform continuously detects the instantaneous jump gradient. When the instantaneous jump gradient falls back into the safe area covered by the environmental physical evolution limit scale and the duration reaches the predetermined test period, the physical boundary anchoring anti-collapse sub-logic is canceled. At the moment of revocation, a smoothing fusion scaling factor is introduced to weight and mix the restored real physical state parameters with the latest derived virtual values ​​according to the smoothing fusion scaling factor, so as to prevent step-fault data from being input into the deep reinforcement learning engine during the recovery period.

[0015] Secondly, the present invention provides an artificial intelligence-based substation electromagnetic interference anti-energy consumption auxiliary control method, comprising the following steps: The physical state parameters of the substation are collected using an array of sensor devices and uploaded to the intelligent control platform via a communication link. Within the intelligent control platform, the collected physical state parameters are input into the internally loaded deep reinforcement learning engine, and the controlled energy-consuming terminal is scheduled according to the action space instructions output by the deep reinforcement learning engine. Continuously monitor the instantaneous gradient of the physical state parameters uploaded by the sensor array; When the instantaneous jump gradient exceeds the pre-calibrated environmental physical evolution limit scale, it is determined that a transient extreme electromagnetic pulse has occurred inside the substation, triggering the physical boundary anchoring anti-collapse sub-logic. When the anti-collapse sub-logic anchored at the physical boundary is in the active phase, it intercepts the distorted physical state parameters and prevents the distorted data from interfering with the input layer of the deep reinforcement learning engine; it retrieves the historical state parameters before the pulse occurs and, combined with the inherent thermodynamic decay constant of the substation building, deduces and generates a virtual slowly changing state sequence, and inputs the virtual slowly changing state sequence into the deep reinforcement learning engine instead of the real parameters. During the activation phase, action permission restrictions are implemented on the output side of the deep reinforcement learning engine. The permission to issue instructions to wake up or start high-power controlled energy-consuming terminals by the deep reinforcement learning engine is cancelled. Only downward control instructions used to maintain the current operating state or reduce the operating load are allowed to prevent dangers caused by secondary power grid surges.

[0016] The present invention has achieved the following beneficial effects: This invention continuously monitors the instantaneous gradient changes of physical state parameters. When an extreme electromagnetic pulse (EMI) is detected within the substation, the anti-collapse sub-logic is activated to promptly intercept distorted environmental parameters, preventing abnormal data containing high-frequency pulse noise from flowing in and interfering with the input layer of the deep reinforcement learning engine. To address the issue of real data being blocked during EMI, the system retrieves historical parameters prior to the pulse occurrence and combines them with the inherent thermodynamic decay constant of the substation structure to generate a virtual, slowly changing state sequence to replace the real parameters input to the control model. This maintains the temporal continuity and physical evolution rationality of the system's input data, preventing logical deadlocks in the algorithm engine due to reception discontinuities. Furthermore, during the anti-collapse sub-logic activation phase, the system implements action permission restrictions on the model's output side, proactively canceling the permission to issue commands to start high-power energy-consuming terminals, retaining only downward control commands to maintain the current state or reduce load. This prevents bus surges caused by erroneous output commands waking up large equipment during interference, effectively reducing the risk of secondary grid surges and improving the operational reliability of the substation's intelligent auxiliary control system in complex electromagnetic environments.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 A diagram illustrating the architecture of an AI-based substation electromagnetic interference anti-electromagnetic interference energy consumption auxiliary control system provided in an embodiment of the present invention. Figure 2 A flowchart illustrating the working principle of the interference level assessment module performing adaptive noise reduction processing provided in this embodiment of the invention; Figure 3 The deep reinforcement learning engine and multi-objective weighted reward structure diagram provided in the embodiments of the present invention; Figure 4 This is a flowchart of the physical boundary anchoring anti-collapse sub-logic activation trigger judgment and control intervention provided in an embodiment of the present invention; Figure 5 A diagram illustrating the anti-crash sub-logic cancellation and smooth hybrid transition mechanism provided in this embodiment of the invention; Figure 6 The flowchart illustrates the overall steps of the AI-based substation electromagnetic interference energy consumption auxiliary control method provided in this embodiment of the invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] This embodiment discloses an artificial intelligence-based substation electromagnetic interference anti-energy consumption auxiliary control system. For example... Figure 1 As shown, the AI-based substation electromagnetic interference anti-energy consumption auxiliary control system includes: a sensor array for collecting physical state parameters of the substation, a controlled energy-consuming terminal for performing environmental adjustment actions, and an intelligent control platform that establishes a communication link with the sensor array and the controlled energy-consuming terminal.

[0022] The substation space houses high-voltage busbars, transformers, and disconnect switches, among other electrical facilities. During equipment operation and switch switching, alternating electromagnetic fields and transient electromagnetic pulses are generated within the substation space. The system described herein is used to perform environmental parameter sensing, data transmission, and energy-assisted control and scheduling under such electromagnetic radiation conditions.

[0023] The sensor array connects to various physical monitoring areas within the substation via communication nodes to acquire real-time environmental status data for both indoor and outdoor environments. The physical status parameters include temperature scale, humidity scale, personnel spatial coordinate point cloud, and light intensity values. The controlled energy-consuming terminal includes various electrical control auxiliary devices within and around the substation, used to adjust the physical environment parameters according to received instructions. The communication link serves as the underlying data transmission channel, transmitting the physical status parameters collected by the sensor array to the intelligent control platform, and transmitting the action space instructions generated by the intelligent control platform to the controlled energy-consuming terminal.

[0024] The intelligent control platform is equipped with a deep reinforcement learning engine. The platform inputs the collected physical state parameters into the deep reinforcement learning engine and schedules the controlled energy-consuming terminals according to the action space instructions output by the engine. The intelligent control platform is built using an edge computing server group. This edge computing server group is deployed in a metal-shielded cabinet in the substation's relay protection room. The grounding terminal of the metal-shielded cabinet is electrically connected to the substation's main grounding grid with low impedance to reduce coupling interference from external electromagnetic fields to the server's motherboard electronic components. The edge computing server group is internally configured with a multi-core central processing unit, a tensor computing array, and non-volatile memory and dynamic random access memory modules.

[0025] The sensor array comprises a temperature detection unit with electromagnetic interference resistance, a presence sensing unit including millimeter-wave radar, and a visual sensing unit equipped with an artificial intelligence image recognition algorithm. The temperature detection unit is located on the monitoring surfaces of the substation control room, relay protection room, and high-voltage switchgear. The external sensing probe and internal circuit board packaging area of ​​the temperature detection unit are covered with a permalloy magnetic shielding layer. The data output pins of the temperature sensing chip inside the temperature detection unit are connected to opto-isolation couplers. The opto-isolation couplers physically separate the front-end sensing circuit and the back-end digital communication circuit on the electrical link, transmitting data through a photoelectric conversion process and blocking the conduction path of common-mode interference current.

[0026] The presence sensing unit is installed in the wall or ceiling structure inside the equipment room and employs a millimeter-wave radar antenna array in a specific ultra-high frequency band. The radar transmitter of the presence sensing unit generates and emits a frequency-modulated continuous wave signal, and the radar receiver captures the echo signal reflected from the human body surface. The digital signal processor inside the presence sensing unit performs a fast Fourier transform on the echo signal, extracting the Doppler frequency shift characteristic signal generated by the rise and fall of the human chest cavity during breathing and heartbeat. Combining this with the phase difference information of the radar antenna array, the presence sensing unit calculates the corresponding spatial coordinate point cloud of the person based on the Doppler frequency shift characteristic signal and continuously outputs the spatial coordinate point cloud of the person as a physical state parameter.

[0027] The visual perception unit includes a high-definition digital camera and a built-in tensor processing computing core. The visual perception unit is installed as a monitoring coverage node in the outdoor area and indoor equipment room of the substation. The tensor processing computing core performs edge local feature extraction and vectorization dimensionality reduction operations on the continuous video stream captured by the camera. The feature extraction process includes inputting the original pixel matrix into a built-in convolutional neural network forward inference model, performing multi-scale feature map mapping, and outputting structured feature vectors representing personnel contours and heat source distribution. The visual perception unit only encapsulates the extracted structured feature data into network data frames and sends them outward, without transmitting the original pixel video data to the intelligent control platform, thereby reducing communication bandwidth consumption.

[0028] The communication link is built using broadband power line carrier technology. The sensor array transmits the physical state parameters, after edge filtering preprocessing, to the intelligent control platform through this communication link. The broadband power line carrier technology reuses existing low-voltage AC power distribution cables in the substation as the physical medium for carrier signal transmission. Each aggregation node of the sensor array is equipped with an orthogonal frequency division multiplexing (OFDM) modem. The OFDM modem converts the serial digital bit stream into a parallel data stream, maps it to constellation nodes, and converts the frequency domain signal into a time domain carrier waveform using a fast inverse Fourier transform, which is then transmitted over the low-voltage AC power distribution cables.

[0029] Before sending the physical state parameters to the orthogonal frequency division multiplexing modem, the edge microcontroller of the sensor array performs edge filtering preprocessing on the raw sampled data stream. This edge filtering preprocessing encompasses a sliding window midpoint filtering algorithm. Within a set time window, the edge microcontroller continuously acquires multiple quantized values ​​as a discrete value sequence; sorts the values ​​in the discrete value sequence arithmetically; and extracts the value at the center of the sorted sequence as the valid physical state parameter for the current processing cycle. The physical state parameters, after edge filtering preprocessing, are packaged into protocol messages and sent to the demodulation receiver on the intelligent control platform side.

[0030] Before the intelligent control platform inputs the physical state parameters into the deep reinforcement learning engine, such as Figure 2 As shown, the intelligent control platform utilizes a built-in interference level assessment module to calculate the statistical variance parameter of the physical state parameters to perform adaptive noise reduction. The interference level assessment module extracts the received physical state parameters, forming a discrete-time parameter sequence. Following a set discrete-time sliding window, the module continuously reads the physical state parameter array within the window, calculates the arithmetic mean of each value in the array, and calculates the statistical variance parameter according to the sum of squared discrete differences formula, used to quantify the degree of dispersion of numerical fluctuations.

[0031] The calculated statistical variance parameter is numerically compared with a pre-defined reference band range. When the statistical variance parameter falls within the pre-defined reference band range, it is determined that the signal currently received by the system contains Gaussian noise whose statistical variance parameter falls within the pre-defined reference band range. Specifically, the pre-defined reference band range is explicitly defined as a closed real-valued interval consisting of the lower and upper limits of the variance. For scalar data such as temperature, the specific interval of this reference band range is calibrated as [0.01, 0.25]. Meanwhile, to ensure the stable reproducibility of the Kalman filter algorithm derivation, the covariance array parameterization includes: the data dimensions of the set process noise covariance matrix, measurement noise covariance matrix, and posterior error covariance matrix are all [0.01, 0.25]. A square matrix of order. During the initial value assignment phase of the system cold start, the process noise covariance matrix is ​​configured such that all diagonal elements are constants. The measurement noise covariance matrix is ​​configured such that all diagonal elements are constants. The diagonal matrix is ​​used, and the initial value of the posterior error covariance matrix is ​​set to a second-order identity matrix to establish the convergence boundary null region. At this time, the interference level assessment module calls the Kalman filter algorithm to perform smoothing calculations on the input data sequence. The execution steps of the Kalman filter algorithm cover the state prediction stage and the observation correction stage. In the state prediction stage, the internally set state transfer matrix is ​​used to substitute the posterior state estimation vector of the previous discrete measurement cycle into matrix multiplication to calculate the prior state prediction vector of the current discrete measurement cycle; simultaneously, based on the posterior error covariance matrix of the previous measurement cycle and the set process noise covariance matrix, the prior error covariance matrix of the current cycle is calculated. During the observation correction phase, the observation data vector pushed by the sensor network at the current moment is extracted. Combined with the observation transfer matrix, the observation residual vector is calculated. Using the prior error covariance matrix and the measurement noise covariance matrix, matrix inversion and multiplication are performed to obtain the Kalman gain matrix. The observation residual vector is multiplied by the Kalman gain matrix and superimposed on the prior state prediction vector to generate the posterior optimal state estimation vector for the current period, which serves as the parameter output after noise reduction. Finally, based on the Kalman gain matrix and the prior error covariance matrix, the posterior error covariance matrix for the current period is refreshed and provided for use in subsequent time steps.

[0032] When the statistical variance parameter exceeds the upper limit of the reference band range, the system is deemed to have encountered electromagnetic interference where the statistical variance parameter exceeds the upper limit of the reference band range. At this time, the interference severity assessment module suspends the Kalman filter algorithm sequence and switches to a composite denoising model configured with a sequence convolutional layer and a bidirectional long short-term memory network layer to perform reconstruction calculations on the interfered physical state parameters. The damaged data sequence within the sliding window is truncated at a fixed time step, converted into a tensor input matrix, and fed into the sequence convolutional layer. The sequence convolutional layer contains multiple discrete one-dimensional convolution sliding operators. These operators slide along the time axis of the time series, performing element-wise multiplication and summation convolution operations on the tensor input matrix within a set receptive field window, combined with set bias parameters and processed by a batch normalization layer. The sequence convolutional layer extracts high-frequency jump components and low-frequency trend components from the original time-domain signal into different feature channels through multiple sets of filter kernels with different parameters, outputting a multi-channel deep feature map tensor.

[0033] The deep feature map tensor is input to the bidirectional long short-term memory (LSTM) network layer. The bidirectional LSM network layer comprises a forward recurrent branch and a backward recurrent branch. The forward recurrent branch reads the deep feature map tensor sequentially according to the direction of time flow, while the backward recurrent branch reads it in the direction of time reversal. In the internal node structure, the forget gate receives the hidden state vector from the previous time step and the input tensor of the current time step, calculates a retention weight matrix with values ​​within a set upper and lower limit range using the logistic function, and performs a Hadamard product operation with the cell state vector from the previous time step. The input gate calculates the candidate feature vector and its corresponding update ratio for the current time step using the logistic function and the hyperbolic tangent function, and accumulates them into the cell state vector to complete the state update. The output gate calculates the hidden layer feature vector for the current time step based on the updated cell state vector and the gating signal. The hidden layer feature vectors output by the forward and backward recurrent branches at their respective time steps are concatenated in the channel dimension. The concatenated feature vector is input into a fully connected mapping layer with a linear activation function, where matrix dimension transformation is performed, and the output is a denoised and reconstructed numerical scalar sequence.

[0034] Specifically, the internal data flow and hyperparameter constraints of the composite noise reduction model are as follows: the input tensor matrix fed into the sequence convolutional layer is fixedly truncated in the time dimension. Each discrete sampling time step (i.e., tensor size is set to [Batch, 64, 1]). The sequence convolutional layer is configured with two consecutively cascaded one-dimensional convolution operators, the first level containing Each kernel has a receptive field window size that is precisely set to... The sliding step size is set to The second level includes One kernel, with a set receptive field window size of [value missing]. The bias parameters attached to the convolution operator all use the mean value as the default bias parameter. variance is The normal distribution constant is initialized. In the bidirectional long short-term memory network layer, the defined upper and lower limit intervals are explicitly defined as the real closed interval [0.0, 1.0] output by the sigmoid activation function; the number of hidden layer state nodes configured inside its forward and reverse recurrent network branches are each fixed at 1. One, extracted after bidirectional splicing. Deep-level integrated feature vectors. The final fully connected mapping layer contains a single layer. Each input neuron and The matrix multiplication operator of the linear output neurons calculates and outputs the current single-step noise reduction value.

[0035] like Figure 3As shown, the deep reinforcement learning engine is built upon a deep Q-network architecture. During normal operation, the deep reinforcement learning engine receives the physical state parameters after interval standardization, the multi-sensor fusion trust index, and the real-time energy consumption of the device as state space elements. The deep Q-network architecture comprises a state input layer, a multi-level fully connected hidden feature extraction layer, and an action value evaluation layer. The neurons within the multi-level fully connected hidden feature extraction layer are constructed using linear rectified activation functions. The action value evaluation layer outputs an action value assessment score representing the expected cumulative benefit of each environmental adjustment action. The action space instructions include control codes for instructing the start / stop status of air conditioning equipment, setting temperature values, adjusting the operating speed of the fresh air exchange system fan, and regulating the duty cycle of intelligent lighting fixtures.

[0036] The interval standardization operation uses an extreme value normalization formula to uniformly map the original parameter values ​​with different dimensions to dimensionless floating-point values. The calculation formula is as follows: ; In the formula, The floating-point value generated for the mapping; The input is the original parameter value; and These are the preset lower limit and upper limit values ​​for parameter measurement, respectively.

[0037] The numerical limit boundary constraint of the dimensionless floating-point array is a closed interval [0.0, 1.0]. To prevent mathematical divergence caused by singular values, the upper and lower limits of various preset parameter measurements are calibrated based on physical hard boundaries: for the ambient temperature parameter, the preset lower limit can be set to -20℃, and the upper limit to 80℃; for the computer room humidity parameter, the lower limit is, for example, 0%RH, and the upper limit is, for example, 100%RH; for the real-time energy consumption of the equipment, the lower limit is, for example, 0.0kW, and the upper limit is set to, for example, 150kW based on the total rated nameplate power constant of all controlled energy-consuming terminals mounted on the substation auxiliary electrical distribution panel.

[0038] The multi-sensor fusion trust index is used to characterize the reliability of the data transmitted back by each sensor node. The intelligent control platform acquires the current sampling set of all sensors of the same type located within the same physical monitoring zone and calculates the statistical average of the sampling set. For an individual sensor node, the absolute value of the difference between the sampled value and the statistical average is calculated. When the absolute value of the difference is less than a preset deviation tolerance threshold, a baseline trust score is assigned to the corresponding sensor node; when the absolute value of the difference is greater than the preset deviation tolerance threshold, or when the packet loss rate of the corresponding communication link is detected to be greater than a preset indicator, a decayed trust score is calculated according to an exponential decay function, and this decayed trust score is the corresponding multi-sensor fusion trust index.

[0039] Specifically, the constant values ​​and algebraic operation rules for the aforementioned criterion include: the benchmark trust score is preferably a constant. Regarding the temperature parameter, the preset deviation tolerance threshold is specified as follows: The preset index for the corresponding underlying packet loss rate of the communication link is defined as a percentage. The specific mathematical algebraic expression of the exponential decay function is defined as follows: .in, For the generated integration trust index; This is the absolute value of the aforementioned difference; The deviation tolerance threshold constant is, for example, ; This is the actual packet loss rate calculated by the underlying polling mechanism. Preset a constant for packet loss metrics, for example, for . The preferred value for the data deviation attenuation constant is [value to be filled in]. ; The communication health penalty attenuation coefficient is set, and the strong penalty, for example, takes a value of [value missing]. This functional form avoids human bias during data cleaning. Specifically, in the formula... Represented by the base of the natural logarithm An exponential function with base , i.e. ; This represents the maximum value function, used to output the largest term among the values ​​within the parentheses. In this formula, as well as The calculation logic is as follows: when the data deviation... or packet loss rate If the difference does not exceed the corresponding set tolerance threshold, the value within the parentheses is negative or zero. The function will take the maximum value. At this point, the corresponding exponential penalty term equals This means that trust levels do not decrease; trust only decreases when the indicator truly exceeds the threshold. The function then outputs the actual over-limit difference and triggers an exponential decay of trust.

[0040] The real-time energy consumption data of the device is read by the energy metering module deployed in the low-voltage switchboard of the substation to obtain the active power consumption data of each controlled energy-consuming terminal, and interval standardization processing is performed simultaneously. The physical state parameters, multi-sensor fusion trust index and real-time energy consumption data of the device after interval standardization are stacked and spliced ​​in the tensor channel dimension to generate a state space tensor for input into the deep Q network architecture.

[0041] The deep reinforcement learning engine employs a reinforcement learning loop that includes an experience replay caching mechanism and updates the policy through a multi-objective weighted reward function. This multi-objective weighted reward function is calculated by aggregating a fusion data accuracy reward component, an energy consumption reduction reward component, and an anti-interference action execution cost penalty component. The calculation mechanism includes: extracting temperature sensor sampling data sequences from different spatial measurement points after an action execution cycle, calculating the spatial distribution variance scalar, and generating the fusion data accuracy reward component based on the inverse mapping of the spatial distribution variance scalar; comparing the cumulative power consumption of the controlled energy-consuming terminal within the current scheduling cycle, and retrieving historical power consumption data from the historical database for the same time period and a similar outdoor temperature scale on a baseline date, calculating the difference between the two (i.e., the energy consumption reduction); querying the action command issuance log, statistically analyzing the command sequences that trigger reverse start / stop for a specific controlled energy-consuming terminal within adjacent consecutive scheduling cycles, and recording the cumulative penalty trigger amount for a single cycle.

[0042] Specifically, the total reward value of a single-step action feedback in a deep reinforcement learning engine ( The following multi-objective weighted reward formula is used to aggregate the rewards: ; In the formula, This is a positive feedback component for data accuracy generated based on spatial distribution variance mapping; This is the difference in energy consumption calculated above (in kilowatt-hours). The specific conversion mapping ratio of the set power conversion constant is preferably a constant of 5.0 (that is, for every 1 kWh decrease in the actual power consumption of the system, the deep Q network is given a bonus of +5.0). The cumulative penalty trigger amount for a single cycle (the set penalty value is added for each individual short-term reverse switching action). The hyperparameter weights are set for each item.

[0043] If the controlled object is a high-power controlled energy-consuming terminal such as a variable frequency air conditioner compressor, in continuous If a start-stop-start or reverse switching action command record is extracted within the minute-by-minute sliding scheduling evaluation window, then a separate value is added for each out-of-bounds reverse switching, for example: The negative penalty constant is assigned to the model to limit the space of dangerous actions.

[0044] Furthermore, The hyperparameter weights are set for each item, and the normalization constraints are satisfied. The specific mapping and switching rules for the dynamic allocation parameters are as follows: Under normal stable operating conditions where the anti-crash sub-logic is not activated, the allocation ratio is set as follows: When the underlying layer detects a strong electromagnetic environmental distortion interruption, the system switches to a conservative defense strategy, resetting to, for example... This means reducing the sensitivity to energy consumption optimization and setting a trigger threshold multiple to enhance the sensitivity to intercepting and penalizing erroneous commands.

[0045] In the training optimization phase, the deep reinforcement learning engine calculates the temporal difference target error based on the state space tensor before and after the action execution, the output action command, and the obtained total reward value. By minimizing the mean squared error loss function, the gradient descent optimization algorithm is used to update the weights and bias node parameters of the deep Q network.

[0046] When high-voltage disconnect switches or circuit breakers in a substation perform opening and closing operations or encounter overvoltage intrusion, they radiate transient extreme electromagnetic pulses into the surrounding space. During this period, the conversion circuits inside the sensor array are affected by the strong electromagnetic field, resulting in distorted physical state parameter data that deviates from the normal range or undergoes steep abrupt changes. If this distorted data with strong impulse noise is fed into the state space nodes of the deep reinforcement learning engine, it will cause a shift in the calculation of activation values ​​in the hidden feature extraction layer, inducing a distorted action probability distribution in the action value evaluation layer. The intelligent control platform is internally deployed with a physical boundary anchoring anti-collapse sub-logic.

[0047] The intelligent control platform continuously monitors the instantaneous gradient changes of the physical state parameters uploaded by the sensor array. The intelligent control platform is equipped with a dedicated computing thread to execute the monitoring tasks.

[0048] The specific mechanism for monitoring the instantaneous gradient jump includes: the system microprocessor uses a high-precision hardware timer to record the arrival timestamps of each physical state parameter data frame. In the data buffer queue, the physical state parameter values ​​and their corresponding timestamps for two adjacent discrete sampling periods of the same monitored object are extracted, and the instantaneous gradient jump (i.e., the physical evolution rate) is calculated using the following formula: ; In the formula, The instantaneous jump gradient is calculated; and These are the physical state parameter values ​​for the current and the previous adjacent discrete sampling periods, respectively. This represents the timestamp span interval corresponding to the two sampling periods mentioned above.

[0049] The calculated instantaneous gradient jump is numerically compared with the maximum possible slope of natural environmental change recorded on the environmental physical evolution limit scale. Changes in the thermodynamic state parameters inside the substation equipment room are limited by physical conditions such as air mass, specific heat capacity physical constant, and the thermal conductivity of the building envelope materials. The rate of change of these macroscopic environmental variables is constrained by objective physical extreme values ​​determined by the aforementioned physical constants. The environmental physical evolution limit scale is stored in the non-volatile memory of the intelligent control platform, and the parameters recorded internally include the maximum possible slope constant of natural environmental change calculated using building heat transfer formulas combined with equipment room volume, wall thermal resistance, and upper limit of heat transfer power.

[0050] like Figure 4 As shown, when the instantaneous jump gradient exceeds the pre-calibrated environmental physical evolution limit scale, the system determines that a transient extreme electromagnetic pulse has occurred inside the substation. The intelligent control platform immediately triggers a system-level interrupt and activates the physical boundary anchoring anti-collapse sub-logic. Specific judgment conditions require multiple cross-domain indicators to be triggered synchronously within the same time evaluation period: the monitoring and calculation unit compares and confirms that the instantaneous jump gradient is greater than the set trigger threshold multiple of the maximum possible environmental natural change slope; simultaneously, the communication interface management module reads the status register of the broadband power line carrier communication underlying protocol stack, obtains the carrier communication packet loss rate and bit error rate statistics within the current evaluation window, and confirms that the carrier communication packet loss rate has risen sharply and exceeded the set communication health warning baseline. When both cross-validation indicators are met simultaneously, the system confirms that strong electromagnetic interference has occurred in the coverage area, determines that a transient extreme electromagnetic pulse has occurred, and activates the physical boundary anchoring anti-collapse sub-logic. The aforementioned multi-dimensional collaborative confirmation mechanism prevents false alarm actions caused by data jumps due to poor local hardware contact.

[0051] The system alarm triggering requires simultaneous determination of two concurrent indicators: the first indicator is a physical abrupt change state, which is triggered when the instantaneous gradient of the real-time environmental parameters reaches 4.0 times or more of the value recorded on the environmental physical evolution limit scale; the second indicator is a communication anomaly state, which is triggered when, within a continuous 100-millisecond communication verification time window, the underlying data packet loss rate is ≥15.0%, or the bit error rate reported by the receiving chip is ≥1×10^(-3). Only when both of the above indicators are met will the system recognize the occurrence of a strong electromagnetic pulse and trigger hardware-level interrupt takeover control.

[0052] While the anti-collapse sub-logic anchored at the physical boundary is active, the intelligent control platform intercepts distorted physical state parameters, preventing distorted data from interfering with the input layer of the deep reinforcement learning engine. The message routing module of the intelligent control platform is configured with data blocking rules. For source address nodes with anomaly flags or those exhibiting instantaneous gradient jumps exceeding limits, the message routing module performs data truncation, stops extracting payload fields, and redirects the message data to the anomaly investigation log storage area, blocking the transmission of state vectors with spike noise to the model assembly module.

[0053] During this cutoff phase, the intelligent control platform retrieves historical state parameters prior to the pulse and, combined with the substation's thermal barrier constant and indoor air specific heat capacity parameters, generates a virtual, slowly changing state sequence, replacing the actual parameters as input to the deep reinforcement learning engine. The specific deduction process involves: retrieving data in reverse order from the historical state queue, skipping distorted sampling frames, and extracting the effective indoor and outdoor air temperatures at their last steady state before the extreme electromagnetic pulse occurred. Simultaneously, it reads the configuration file to obtain the thermal barrier constant (derived by dividing the wall material thickness by the thermal conductivity coefficient) and the indoor air specific heat capacity parameter (derived by multiplying the computer room's net volume by the air density at a specific temperature and the constant-pressure specific heat capacity).

[0054] Substituting the thermal barrier constant, the specific heat capacity of indoor air, the effective heat exchange surface area of ​​the substation building, and the temperature difference between indoors and outdoors into a preset thermodynamic hysteresis decay formula, the natural temperature drift step size under the premise of no active control intervention from any HVAC equipment is calculated. The simulation engine uses the Fourier heat transfer discrete difference formula to calculate the natural temperature drift step size for a single period. ): ; In the formula, and These correspond to the effective indoor and effective outdoor temperatures recorded in the last steady-state state before the pulse, respectively. To deduce the infinitesimal time slice (i.e., the discrete-time infinitesimal step size parameter), it is set to be equal to the actual polling sampling period of the system, and the value is, for example, a constant of 1.0 seconds; The thermal resistance constant of the substation main body is determined based on the constantized thermal conductivity of the general-purpose color steel sandwich core material in the equipment room, for example, as follows: ; The effective surface area parameter for heat exchange between the main body of the substation building and the external environment is calculated based on the dimensions of the external envelope of the equipment room building, and is, for example, a scalar constant. ; The specific heat capacity of the indoor air is calculated based on the three-dimensional net volume of the substation and the air density, and its value is, for example, a scalar constant. In subsequent processing sequences, the algorithm performs high-frequency scalar addition and superposition using this constant step size difference value to simulate the extension of the real thermodynamic hysteresis curve.

[0055] The intelligent control platform continuously accumulates the temperature natural drift step size over time to generate a coherent virtual slowly varying state sequence. Specifically, using the last effective indoor temperature before the pulse as the initial baseline, in each projection cycle, the predicted temperature value of the previous cycle is added to the temperature natural drift step size of the current cycle to generate a new predicted temperature. Simultaneously, in subsequent cycles, the indoor-outdoor temperature difference is recalculated to update the drift step size, and this iterative accumulation continues. Finally, the platform formats the generated virtual slowly varying state sequence into a tensor structure supported by a neural network, replacing the distorted real parameters as input to maintain the temporal continuity of the decision network input and prevent system collapse due to data gaps.

[0056] During the activation phase, the intelligent control platform imposes action permission restrictions on the output side of the deep reinforcement learning engine, canceling the deep reinforcement learning engine's permission to issue commands to wake up or start high-power controlled energy-consuming terminals, and only allowing downward control commands to maintain the current operating state or reduce the operating load, in order to prevent secondary power grid surges induced by the accidental start-up of high-power equipment during extreme electromagnetic pulses. High-power electrical auxiliary equipment will absorb a large start-up surge from the power supply bus during the cold start phase. The specific execution logic of the action permission restrictions includes: deploying command mask filtering probes in the command output buffer layer of the intelligent control platform. The command mask filtering probes perform feature bit matching verification on the control messages to be encapsulated and sent to the lower layer. When the instruction control field contains hexadecimal codes indicating load increase features such as closing an AC contactor to open equipment, increasing the AC output frequency of the frequency converter, or connecting a high-power electric heating wire, the instruction mask filtering probe intercepts and discards the target message; when the control field contains hexadecimal codes indicating load reduction features such as maintaining the existing switch operation state without action, reducing the operating frequency, or performing a power-off shutdown operation, the instruction mask filtering probe allows the target message to pass and sends it to the associated controlled energy-consuming terminal through the underlying communication interface.

[0057] The instruction mask filtering probe performs filtering operations based on a preset mapped list of underlying hexadecimal action opcodes. The logical analysis and masking extraction rules are as follows: capture the TCP / IP payload field of the underlying packet... If the control command of a byte matches a hexadecimal code belonging to a surge-inducing load-increasing characteristic array cluster (e.g., representing main circuit contactor closure: 0x1A, driving electric heating resistance wire energized: 0x2B, fan frequency conversion frequency increase: 0x3C), it is directly intercepted and discarded. When the detected control field command code belongs to a load-reducing operation characteristic array (e.g., shutdown hibernation action: 0x00, maintaining the current state suspension: 0xFF, or cutting off compressor cooling and only supplying air: 0xA1), the judgment rule outputs a security verification pass signal, which is directly allowed through the network card protocol stack.

[0058] As the energy of transient extreme electromagnetic pulse events dissipates, such as Figure 5 As shown, the intelligent control platform continuously detects the instantaneous jump gradient. When the monitoring and calculation unit confirms that within a continuously set set of evaluation sampling periods, the instantaneous jump gradient has fallen back to the safe parameter area covered by the environmental physical evolution limit scale, and the broadband power line carrier communication packet loss rate has recovered to the set normal operating range, and the duration of the aforementioned dual stability parameters has reached the set predetermined test period, the system main control thread releases the interrupt lock state and cancels the activation instruction of the physical boundary anchoring anti-crash sub-logic.

[0059] During the revocation transition phase, the system re-establishes the route for reading the real physical state parameters transmitted back from the sensor array. To address the numerical shift difference between the purely mathematically derived data sequence and the re-accessed real sensor data sequence, a smoothing fusion scaling factor is introduced. The restored real physical state parameters and the latest derived virtual values ​​are weighted and mixed according to this smoothing fusion scaling factor to prevent step-discontinuity data from being input into the deep reinforcement learning engine during the recovery period. The smoothing fusion scaling factor is configured as a floating-point coefficient variable that increases progressively with each processing time step. When the smooth transition time window is initiated, the mixed state input values ​​for each individual time step are calculated according to the following weighted mixing algebraic equation: ; In the formula, Input values ​​for the generated mixed state; To restore the physical characteristic values ​​transmitted from the actual acquisition end; These are virtual feature values ​​derived by the concurrent algorithm. Constant factors in the formula. For example, it can be configured as a constant of 1.0; the smoothing fusion scaling factor The specific arithmetic progression function in the undo phase is defined as follows: .in, This represents the accumulated clock steps taken since the start of the recovery process within the set time window. The total length scalar encompassed by the set transition time window is preferably fixed at 30 discrete operation steps (equivalent to a 30-second transition interval). During this process... During the transition period, the data rises gently in a sloping curve and is capped at the saturation extreme value of 1.0, thus completing the smooth transition and handover of the data.

[0060] When the set transition time window is exhausted, the smooth fusion scaling factor increases to the saturation limit, the system stops the virtual numerical derivation and weighted hybrid operation branches, and directly inputs the pure real physical state parameters into the deep reinforcement learning engine to achieve a smooth handover and recovery of the control loop.

[0061] Regarding the specific structure and control actions of the controlled energy-consuming terminal, the controlled energy-consuming terminal includes an intelligent lighting array. The intelligent control platform, based on the feature data obtained from structured parsing of the video stream uploaded by the visual perception unit, runs a target detection model to determine whether there is an unauthorized intrusion or fire incident within the substation. The intelligent lighting array is distributed in a grid matrix on the outdoor area of ​​the substation and the top of the indoor relay protection room. Each lamp contains a light-emitting diode module, a stepless dimming drive circuit, and a digital node module supporting carrier-addressed communication. The target detection model is constructed using a convolutional neural network operator architecture including residual densely connected channels and a feature pyramid network. It receives structured feature data and performs candidate anchor box regression calculations. Through comparative calculations, the system identifies the geometric features of the outline of personnel not wearing insulated reflective safety clothing, as well as the abnormally bright pixel feature areas of the initial arc discharge caused by short-circuit breakdown of high-voltage equipment.

[0062] Specifically, the steps for determining intrusion by personnel not wearing insulated reflective clothing include: First, converting the human image region extracted from the boundary candidate anchor frame from the RGB color gamut to the HSV color gamut; second, selecting pixels that simultaneously satisfy the following conditions: hue (H) parameter in the closed interval [35, 60], saturation (S) parameter ≥ 0.6, and lightness (V) parameter in the range [0.7, 1.0], as the reflective protective gear pixel cluster. When the area of ​​the reflective protective gear pixel cluster accounts for less than the proportion of the total pixel area of ​​the human body, the system outputs a violation alarm signal for not wearing standard protective clothing.

[0063] The specific logical judgment conditions and steps for identifying the initial arc discharge fire are as follows: A continuous time interval is extracted using a visual module, for example... A single-channel grayscale video tracking frame stream with a step size of one second; a morphological thresholding algorithm is used to extract and separate grayscale intensity values ​​in the first frame data stream. Physically connected bright clusters of pixels form pixel clusters; during the subsequent three-frame tracking period, a cluster's two-dimensional plane centroid (X, Y coordinates) shifts by a certain amount during the tracking period if and only if the following conditions are met: Each spatial pixel unit span (determined to possess fixed discharge characteristics, eliminating headlight shadow interference), and the single-frame ratio expansion of the bounding box of this pixel cluster increases dramatically. When the expansion characteristics of deflagration plasma are established, the composite judgment result takes effect, and the occurrence of an electrical breakdown arc fire is accurately determined.

[0064] Furthermore, the system's internal logic determines that the aforementioned pre-calibrated physical lower limit constant for the minimum heat exchange temperature difference threshold is strictly set to... When the physical difference in heat transfer is not reached, the enthalpy reduction benefit from the work done by the forced intake airflow will be completely offset by the electrical heating generated by the armature of the centrifugal fan, resulting in ineffective reverse feedback energy consumption. When the condition is met... The difference threshold is a prerequisite for the underlying permissions to execute the fresh air action.

[0065] When an unauthorized intrusion or fire incident is detected, the intelligent control platform outputs an emergency dimming command, triggering the intelligent lighting array to activate its maximum rated brightness illumination mode and simultaneously sending an alarm message to the external security control center. The emergency dimming command embeds a high-priority preemption identifier. Upon receiving the command, the communication node of the intelligent lighting array rapidly adjusts the duty cycle of its output drive current signal to the hardware's rated full load state using its internal pulse width modulation controller. Simultaneously, the intelligent control platform generates an event log, including trigger timestamp parameters, substation digital model location coordinates, and anomaly capture image matrix data stored by the visual perception unit. This data is encapsulated and synthesized into the alarm message using a high-level encryption algorithm and pushed to the server receiving port of the external security control center via the substation's dedicated fiber optic backbone ring network.

[0066] The controlled energy-consuming terminals include intelligent air conditioning units, electric heaters, and a fresh air exchange system; the intelligent control platform is embedded with indoor and outdoor temperature linkage comparison logic. The secondary microcomputer electrical control equipment inside the substation continuously generates Joule heat during operation. Under specific meteorological conditions or at night, the outdoor atmospheric temperature is relatively lower than the indoor temperature. The indoor and outdoor temperature linkage comparison logic is configured as follows: continuously reading the outdoor temperature value from the outdoor temperature detection unit and the expected target set temperature value indoors; calculating the difference between the two, determining that the outdoor temperature value is lower than the indoor target set temperature and that the temperature difference gradient between the two exceeds a pre-calibrated minimum heat exchange temperature difference threshold.

[0067] When the outdoor ambient temperature is determined to meet the substation's cooling requirements and the movement trajectories of indoor personnel are within a safe isolation zone, the intelligent control platform prioritizes issuing a command to activate the fresh air exchange system to introduce outdoor cold air. The system queries the personnel spatial coordinate point cloud tracking data provided by the presence sensing unit and compares it with the three-dimensional space inside the equipment room to delineate a reference area. When it is confirmed that there are no personnel present in the current building space, or that the movement spatial coordinates of all personnel are within the windproof safe isolation zone outside the physical blowing area of ​​the fresh air system, the intelligent control platform issues a start command to the fresh air exchange system. After receiving the action command, the drive controller of the fresh air exchange system drives the stepper motor to adjust the rotation angle of the dustproof air inlet and outlet louver guide vanes, and starts the high-power centrifugal fan to forcibly draw low-enthalpy outdoor cold air into the indoor space. Through air pressure convection, the hot air inside is expelled outside, achieving indoor cooling through mechanical physical ventilation. After a set working time interval, the slope of the indoor temperature scale change is extracted. When it is determined that the introduction of outdoor cold air source cannot achieve the cooling target, the intelligent control platform switches the control strategy branch, sends an instruction message to the intelligent air conditioning unit, starts the air conditioning inverter compressor and adjusts the throttle valve to perform the reverse Carnot cycle supplementary cooling task.

[0068] Based on the aforementioned system module construction and underlying data flow interaction logic, this embodiment also discloses an artificial intelligence-based substation electromagnetic interference anti-energy consumption auxiliary control method. For example... Figure 6 As shown, this method relies on the aforementioned AI-based substation electromagnetic interference anti-energy consumption auxiliary control system for execution, and the specific steps include: Step S10: Collect physical state parameters of the substation using a sensor array and upload these parameters to the intelligent control platform via a communication link. Temperature detection units, presence sensing units, and visual sensing units deployed in various areas of the substation continuously sample environmental indicators using built-in analog-to-digital conversion circuits, converting them into serial digital signal frames. The sensor edge processing node preprocesses the acquired data sequence using a sliding window mid-value filtering algorithm; subsequently, it modulates the digital signal frames into orthogonal frequency division multiplexing carrier signals and transmits them to the edge computing aggregation node via a low-voltage AC power distribution cable carrier channel.

[0069] Step S20: Within the intelligent control platform, the built-in interference level assessment module extracts parameter sequences from the received physical state parameters to calculate statistical variance parameters for adaptive noise reduction processing. When the statistical variance parameter is within a pre-set reference band range, the Kalman filter algorithm is scheduled to perform state prediction and observation correction smoothing processing on the data sequence based on the state transfer matrix; when the statistical variance parameter exceeds the upper limit of the reference band range, a composite noise reduction model assembled from sequence convolutional layers and bidirectional long short-term memory network layers is invoked to perform tensor extraction and depth reconstruction operations on the environmental data feature sequences containing high-frequency jump distortions.

[0070] Step S30: The collected and denoised physical state parameters are input into the internally loaded deep reinforcement learning engine, and the controlled energy-consuming terminals are scheduled according to the action space instructions output by the deep reinforcement learning engine. The physical state parameters, multi-sensor fusion trust index, and real-time energy consumption values ​​of the devices are subjected to interval standardization numerical mapping, spliced ​​and stacked into a multi-dimensional input tensor, and input into the deep Q-network architecture; gradient optimization is performed using the instantaneous scalar reward value generated by aggregating the fusion data accuracy reward component, the energy consumption reduction reward component, and the anti-interference action execution cost penalty component, and the specific instruction parameter message for controlling each controlled energy-consuming terminal is output.

[0071] Step S40: Continuously monitor the instantaneous gradient of the physical state parameters uploaded by the sensor array. Allocate an independent background computing thread to extract the numerical difference of the physical state parameters between two adjacent discrete sampling periods at hardware clock polling intervals. Divide the numerical difference by the time span scalar of the adjacent sampling periods to calculate the current gradient slope index.

[0072] Step S50: When the instantaneous jump gradient exceeds the pre-calibrated environmental physical evolution limit scale, a transient extreme electromagnetic pulse is determined to have occurred inside the substation, triggering the physical boundary anchoring anti-collapse sub-logic. The comparison and calculation module performs a numerical comparison and determines that the difference between the current jump slope and the set maximum possible environmental natural change slope constant exceeds the trigger warning line. When the broadband power line carrier communication packet loss rate calculated by the network underlying protocol stack meets the incremental over-limit triggering condition, a transient strong interference event state is established, an interrupt is thrown, and the sub-logic action is triggered.

[0073] Step S60: While the anti-collapse sub-logic anchored at the physical boundary is active, distorted physical state parameters are intercepted to prevent distorted data from interfering with the input layer of the deep reinforcement learning engine. Address masking and feature interception masking rules are enabled at the platform data stream processing queue level to redirect message sequences containing abnormal transition slope identifiers to discarded cache addresses for processing.

[0074] Step S70: Retrieve historical state parameters before the pulse occurs and combine them with the inherent thermodynamic decay constant of the substation building to generate a virtual slowly changing state sequence. Input the virtual slowly changing state sequence into the deep reinforcement learning engine, replacing the actual parameters. Extract the last recorded effective indoor and outdoor temperature values ​​in a stable band before the communication anomaly interruption event. Utilize the set building thermal barrier constant and indoor air specific heat capacity parameters, substitute them into Newton's cooling heat flux formula to calculate the dissipation rate, and solve for the temperature natural drift step size for each discrete period. Continuously accumulate the temperature natural drift step size on the time extrapolation axis, and output a coherent and smooth virtual slowly changing state sequence to the deep reinforcement learning engine node.

[0075] Step S80: During the activation phase, action permission restrictions are implemented on the output side of the deep reinforcement learning engine. The permission to issue instructions to wake up or start high-power controlled energy-consuming terminals by the deep reinforcement learning engine is revoked. Only downward control instructions used to maintain the current operating state or reduce the operating load are allowed to prevent dangers caused by secondary grid surges. A message output register interception filter is configured to parse and match opcodes and intercept data packets containing active power increment command frames in the control word. Only data packets that maintain the current power operation or perform shutdown attenuation derating are approved for transmission to the underlying link.

[0076] Step S90: When the instantaneous gradient jump is detected to have fallen back into the safe region covered by the environmental physical evolution limit scale and the duration reaches the predetermined test period, the physical boundary anchoring anti-collapse sub-logic is revoked. During the transition phase of the unblocking procedure, the real physical state parameters restored to normal and the latest derived virtual values ​​are proportionally weighted and mixed according to the set smoothing fusion scaling factor that increases with the computation time step. The mixed numerical sequence containing transition smoothing weights is input into the deep reinforcement learning engine. After the transition window ends, the control blockade is lifted, and the normal communication control loop of the system is restored.

[0077] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An AI-based substation electromagnetic interference anti-energy consumption auxiliary control system, comprising a sensor array for collecting physical state parameters of the substation, a controlled energy-consuming terminal for performing environmental adjustment actions, and an intelligent control platform that establishes a communication link with the sensor array and the controlled energy-consuming terminal; The intelligent control platform is equipped with a deep reinforcement learning engine. The intelligent control platform is used to input the collected physical state parameters into the deep reinforcement learning engine and to schedule the controlled energy-consuming terminal according to the action space instructions output by the deep reinforcement learning engine. Its features are, The intelligent control platform is internally deployed with physical boundary anchoring anti-crash sub-logic. The intelligent control platform continuously monitors the instantaneous gradient of the physical state parameters uploaded by the sensor array; When the instantaneous jump gradient exceeds the pre-calibrated environmental physical evolution limit scale, it is determined that a transient extreme electromagnetic pulse has occurred inside the substation, and the intelligent control platform triggers the physical boundary anchoring anti-collapse sub-logic. When the anti-collapse sub-logic anchored at the physical boundary is in the active phase, the intelligent control platform intercepts distorted physical state parameters and prevents distorted data from interfering with the input layer of the deep reinforcement learning engine. At the same time, the intelligent control platform retrieves historical state parameters before the pulse occurs and, in conjunction with the inherent thermodynamic decay constant of the substation building, deduces and generates a virtual slowly changing state sequence. The virtual slowly changing state sequence is then used to replace the real parameters and input into the deep reinforcement learning engine. Furthermore, during the activation phase, the intelligent control platform restricts the action permissions of the output side of the deep reinforcement learning engine, cancels the permission of the deep reinforcement learning engine to wake up or start high-power controlled energy-consuming terminals, and only allows downward control commands to maintain the current operating state or reduce the operating load, so as to prevent secondary power grid surges caused by the accidental start of high-power equipment during extreme electromagnetic pulses.

2. The substation electromagnetic interference anti-electromagnetic interference energy consumption auxiliary control system based on artificial intelligence as described in claim 1, characterized in that, The sensor array includes a temperature detection unit with anti-electromagnetic interference characteristics, a presence sensing unit containing millimeter-wave radar, and a visual sensing unit equipped with artificial intelligence image recognition algorithms. The communication link is built on broadband power line carrier technology, and the sensor array transmits the physical state parameters after edge filtering preprocessing to the intelligent control platform through the communication link.

3. The substation electromagnetic interference anti-electromagnetic interference energy consumption auxiliary control system based on artificial intelligence as described in claim 1, characterized in that, The deep reinforcement learning engine is built upon a deep Q-network architecture. During normal operation, the deep reinforcement learning engine receives the physical state parameters after interval standardization, the multi-sensor fusion trust index, and the real-time energy consumption value of the device as state space elements. The deep reinforcement learning engine optimizes policies through a multi-objective weighted reward function, which is composed of a fusion data accuracy reward component, an energy consumption reduction reward component, and an anti-interference action execution cost penalty component aggregated according to dynamic matching parameters.

4. The substation electromagnetic interference anti-electromagnetic interference energy consumption auxiliary control system based on artificial intelligence as described in claim 2, characterized in that, The controlled energy-consuming terminal includes an array of intelligent lighting fixtures; The intelligent control platform runs a target detection model based on the video stream uploaded by the visual perception unit to determine whether there is an unauthorized intrusion or fire incident inside the substation. When an unauthorized intrusion or fire is detected, the intelligent control platform outputs an emergency dimming command, triggering the intelligent lighting array to activate its maximum rated brightness illumination mode, and simultaneously pushes an alarm message to the external security control center.

5. The substation electromagnetic interference anti-energy consumption auxiliary control system based on artificial intelligence according to claim 1, characterized in that, The controlled energy-consuming terminals include intelligent air conditioning units, electric heating devices, and fresh air exchange systems; The intelligent control platform has an embedded logic for comparing indoor and outdoor temperatures. When it is determined that the outdoor natural temperature meets the cooling requirements of the substation and the activity trajectory of the indoor staff is within a safe isolation zone, the intelligent control platform will prioritize issuing an instruction to start the fresh air exchange system to introduce outdoor cold air. When it is determined that introducing outdoor cold air source cannot achieve the cooling target, the intelligent control platform switches control commands to start the intelligent air conditioning unit to perform supplementary cooling tasks.

6. The substation electromagnetic interference anti-electromagnetic interference energy consumption auxiliary control system based on artificial intelligence according to claim 1, characterized in that, Before inputting the physical state parameters into the deep reinforcement learning engine, the intelligent control platform uses a built-in interference level assessment module to calculate the statistical variance parameter of the physical state parameters in order to perform adaptive noise reduction. When Gaussian noise is detected within the preset reference band range for the statistical variance parameter, the interference level assessment module schedules the Kalman filter algorithm to smooth the data sequence. When electromagnetic interference is detected where the statistical variance parameter exceeds the upper limit of the reference band range, the interference degree assessment module switches to and calls a composite noise reduction model assembled from sequential convolutional layers and bidirectional long short-term memory network layers to perform deep reconstruction of the damaged environmental data features.

7. The substation electromagnetic interference anti-electromagnetic interference energy consumption auxiliary control system based on artificial intelligence according to claim 2, characterized in that, The specific action mechanism by which the intelligent control platform monitors the instantaneous gradient jump is as follows: The numerical difference of the physical state parameter between two adjacent sampling periods is extracted at the microsecond level, and the numerical difference is divided by the sampling period duration to obtain the current jump slope. Compare the current jump slope with the maximum possible slope of natural environmental change recorded on the environmental physical evolution limit scale. When the current transition slope reaches a multiple of the set trigger threshold of the maximum possible environmental natural change slope, and at the same time the packet loss rate of the underlying broadband power line carrier communication increases and exceeds the set communication health warning baseline, the occurrence of the transient extreme electromagnetic pulse is finally confirmed.

8. The substation electromagnetic interference anti-energy consumption auxiliary control system based on artificial intelligence according to claim 1, characterized in that, The specific mechanism for generating the virtual slowly changing state sequence through deduction is as follows: Extract the last recorded effective indoor temperature value and effective outdoor temperature value before the transient extreme electromagnetic pulse occurs, and obtain the thermal resistance constant of the outer wall of the substation building, the effective heat exchange surface area parameter of the substation building, and the indoor air specific heat capacity parameter. Substituting the thermal barrier constant, the effective heat exchange surface area parameter, the indoor air specific heat capacity parameter, and the indoor-outdoor temperature difference into the preset thermodynamic hysteresis decay formula, the temperature natural drift step size under the condition of losing environmental control measures is calculated. Based on the temperature natural drift step size, the virtual slowly changing state sequence is accumulated on the time axis.

9. The substation anti-electromagnetic interference energy consumption auxiliary control system based on artificial intelligence according to claim 1, characterized in that, The intelligent control platform continuously detects the instantaneous jump gradient. When the instantaneous jump gradient falls back into the safe area covered by the environmental physical evolution limit scale and the duration reaches the predetermined test period, the physical boundary anchoring anti-collapse sub-logic is canceled. At the moment of revocation, a smoothing fusion scaling factor is introduced to weight and mix the restored real physical state parameters with the latest derived virtual values ​​according to the smoothing fusion scaling factor, so as to prevent step-fault data from being input into the deep reinforcement learning engine during the recovery period.

10. A substation electromagnetic interference anti-energy consumption auxiliary control method based on artificial intelligence, characterized in that, Includes the following steps: The physical state parameters of the substation are collected using an array of sensor devices and uploaded to the intelligent control platform via a communication link. Within the intelligent control platform, the collected physical state parameters are input into the internally loaded deep reinforcement learning engine, and the controlled energy-consuming terminal is scheduled according to the action space instructions output by the deep reinforcement learning engine. Continuously monitor the instantaneous gradient of the physical state parameters uploaded by the sensor array; When the instantaneous jump gradient exceeds the pre-calibrated environmental physical evolution limit scale, it is determined that a transient extreme electromagnetic pulse has occurred inside the substation, triggering the physical boundary anchoring anti-collapse sub-logic. When the anti-collapse sub-logic anchored at the physical boundary is in the active phase, it intercepts the distorted physical state parameters and prevents the distorted data from interfering with the input layer of the deep reinforcement learning engine; it retrieves the historical state parameters before the pulse occurs and, combined with the inherent thermodynamic decay constant of the substation building, deduces and generates a virtual slowly changing state sequence, and inputs the virtual slowly changing state sequence into the deep reinforcement learning engine instead of the real parameters. During the activation phase, action permission restrictions are implemented on the output side of the deep reinforcement learning engine. The permission to issue instructions to wake up or start high-power controlled energy-consuming terminals by the deep reinforcement learning engine is cancelled. Only downward control instructions used to maintain the current operating state or reduce the operating load are allowed to prevent dangers caused by secondary power grid surges.

Images