Management system and device combining artificial intelligence and wind and light storage application

Abstract

The invention relates to the technical field of automatic control and intelligent operation and maintenance of a power system, in particular to a management system and device combining artificial intelligence and wind and light storage application. The multi-dimensional fusion sensing module is configured to respectively acquire electrical operation parameters, environment temperature and humidity data and monitoring video frames of load equipment by utilizing respective independent preset sampling clocks, execute time sequence alignment operation on the electrical operation parameters, the environment temperature and humidity data and the monitoring video frames, and output multi-dimensional fusion sensing records; and based on the multi-dimensional fusion perception record, extracting environment state features in the multi-dimensional fusion perception record. According to the method, the real-time electrical parameters and the monitoring video image features are deeply fused, and the weighted Euclidean distance and the convolutional neural network are utilized to realize multi-dimensional state perception, so that the defect that the traditional single threshold monitoring is easily interfered by the environment and misinformation is caused is effectively overcome, and the accuracy of fault identification of the wind-solar storage system is improved.

Description

A management system and device integrating artificial intelligence with wind, solar and energy storage applications Technical Field

[0001] This invention relates to the field of power system automation control and intelligent operation and maintenance technology, and in particular to a management system and device that combines artificial intelligence with wind, solar and energy storage applications. Background Technology

[0002] Driven by the "dual carbon" goal, integrated microgrid systems combining distributed wind power, photovoltaic power generation, and electrochemical energy storage have become a key support for building new power systems. These systems are typically deployed in remote areas or industrial parks with complex natural environments, involving the coordinated operation of numerous heterogeneous devices such as photovoltaic modules, wind turbine converters, energy storage BMS, and grid-connected inverters. To ensure the stability and security of power supply, modern energy management systems have widely adopted multi-source sensing technology, utilizing high-precision sensors to collect real-time electrical analog quantities such as voltage, current, and frequency, and deploying high-definition monitoring cameras to monitor the site environment around the clock. Simultaneously, with the penetration of Industry 4.0 and artificial intelligence technologies, digital twin technology is beginning to be applied to the operation and maintenance of wind, solar, and energy storage systems. By building virtual mapping models of physical equipment at the edge or in the cloud, real-time simulation and historical backtracking of the power generation equipment's operating status can be achieved, aiming to replace the traditional manual inspection mode and improve the intelligent operation and maintenance level of new energy power stations.

[0003] However, existing wind, solar and energy storage management systems rely heavily on static thresholds for single electrical parameters for fault diagnosis. When faced with complex dynamic conditions such as photovoltaic cloud shading, wind turbine gust disturbances, or frequent switching of energy storage, they are unable to effectively distinguish between normal transient fluctuations and real equipment failures, resulting in a high false alarm rate. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a management system and device that combines artificial intelligence with wind, solar and energy storage applications, aiming to improve the problem that existing wind, solar and energy storage management systems rely heavily on static threshold judgment of a single electrical parameter for fault diagnosis.

[0005] In a first aspect, the present invention provides the following technical solution: a management system combining artificial intelligence with wind, solar, and energy storage applications, comprising:

[0006] The multi-source data acquisition module is configured to acquire electrical operating parameters of the load device, ambient temperature and humidity data, and monitoring video frames using their respective independent preset sampling clocks, perform time alignment operations on the electrical operating parameters, ambient temperature and humidity data, and monitoring video frames, and output multi-dimensional fusion perception records.

[0007] The edge digital twin module extracts environmental state features based on the multi-dimensional fusion perception record, retrieves the theoretical benchmark operation feature vector corresponding to the environmental state features using a preset benchmark operation model, calculates the numerical deviation between the electrical operation parameters and the theoretical benchmark operation feature vector, maps the numerical deviation to a preset business state space, and outputs business state classification labels.

[0008] The power supply strategy mapping module, based on the service status classification label, retrieves the target power supply mode corresponding to the service status classification label from the pre-stored status-action mapping table, and generates a dual-path switching combination instruction that includes the target conduction status of the mains circuit breaker and the target conduction status of the energy storage circuit breaker.

[0009] The physical link execution module, based on the dual-path switching combination command, drives the mains circuit breaker and the energy storage circuit breaker to operate to the target conduction state defined by the command, collects the contact signals after the circuit breaker operates, and outputs the current topology status feedback.

[0010] The model adaptive correction module, based on the current topology state feedback and the historical deviation data formed by the accumulation of numerical deviations, calculates the parameter weight adjustment value of the benchmark running model through an error feedback-based model correction algorithm, and writes the parameter weight adjustment value as the model update coefficient into the edge digital twin module.

[0011] Preferably, in the multi-source data acquisition module, the step of performing time-series alignment operations on the electrical operating parameters, the ambient temperature and humidity data, and the monitoring video frames to output a multi-dimensional fused sensing record includes:

[0012] The received electrical operating parameters, ambient temperature and humidity data, and monitoring video frames are marked with a timestamp using a preset unified system clock.

[0013] Extract the frame generation timestamp of the monitoring video frame and set the frame generation timestamp as the reference alignment anchor point for data fusion;

[0014] In the time series of the electrical operating parameters, two adjacent sampling points located before and after the reference alignment anchor point are retrieved, and the electrical fitting value of the electrical operating parameters at the reference alignment anchor point time is calculated using a linear interpolation algorithm;

[0015] In the time series of the environmental temperature and humidity data, a sampling point whose absolute value of the time difference with the benchmark alignment anchor point is lower than that of other sampling points is locked as the environmental retention value.

[0016] The image encoding data of the monitoring video frame, the electrical fitting value, and the environmental maintenance value are encapsulated in a unified protocol format to generate the multi-dimensional fusion sensing record.

[0017] Preferably, in the edge digital twin module, the step of calculating the numerical deviation between the electrical operating parameters and the theoretical reference operating feature vector includes:

[0018] The multi-dimensional fusion sensing record is analyzed to separate the real-time electrical feature vector and environmental state features;

[0019] The environmental state features are input into the multi-dimensional state mapping unit in the benchmark operation model to retrieve the theoretical benchmark operation feature vector corresponding to the environmental state features.

[0020] For the corresponding dimensions of the real-time electrical feature vector and the theoretical benchmark operating feature vector, calculate the absolute value of the component difference to generate a difference vector.

[0021] A weighted Euclidean distance operation is performed on the difference vector, and the scalar result obtained from the operation is determined as the numerical deviation.

[0022] Preferably, in the edge digital twin module, the step of mapping the numerical deviation to a preset business state space and outputting business state classification labels includes:

[0023] Extract the monitoring video frames from the multi-dimensional fusion perception record, input the monitoring video frames into a preset image anomaly recognition neural network, and extract the visual anomaly confidence value.

[0024] Construct a joint state feature set that includes the numerical deviation and the confidence value of the visual anomaly;

[0025] The joint state feature set is input into a pre-trained Softmax classifier to calculate and generate a probability distribution vector corresponding to each sub-state in the preset business state set.

[0026] Extract the state index corresponding to the component whose value is higher than the other components in the probability distribution vector, and convert the state index into a business state classification label in text format.

[0027] Preferably, in the power supply strategy mapping module, the step of generating a dual-path switching combination command that includes the target conduction state of the mains circuit breaker and the target conduction state of the energy storage circuit breaker includes:

[0028] Extract the index key value of the business status category label;

[0029] Using the index key value as the retrieval keyword, locate the target control strategy row in the state-action mapping table, and read the preset mains circuit control identifier and energy storage circuit control identifier in the target control strategy row;

[0030] Construct an instruction data packet containing a frame header sequence, device physical address, and function code;

[0031] Write the mains circuit control identifier into the mains control bit of the instruction data packet, write the energy storage circuit control identifier into the energy storage control bit of the instruction data packet, and generate a dual-path switching combination instruction.

[0032] Preferably, in the physical link execution module, the step of driving the mains circuit breaker and the energy storage circuit breaker to operate to the target conduction state defined by the instruction includes:

[0033] The dual-channel switching combination command is analyzed to separate the mains control bit and energy storage control bit contained therein;

[0034] The separated mains control bits are converted into mains drive voltage signals, and the separated energy storage control bits are converted into energy storage drive voltage signals.

[0035] The mains drive voltage signal is applied to the operating coil of the mains circuit breaker through the first isolation drive channel, driving the moving contact of the mains circuit breaker to move to the mains target position corresponding to the target conduction state of the mains circuit breaker.

[0036] Simultaneously, the energy storage drive voltage signal is applied to the operating coil of the energy storage circuit breaker through the second isolation drive channel, driving the moving contact of the energy storage circuit breaker to move to the energy storage target position corresponding to the target conduction state of the energy storage circuit breaker.

[0037] Preferably, in the physical link execution module, the step of outputting the current topology status feedback includes:

[0038] Collect the first potential signal of the auxiliary contacts of the mains circuit breaker and the second potential signal of the auxiliary contacts of the energy storage circuit breaker;

[0039] The first potential signal is converted into an actual status code of the mains power using a preset signal acquisition circuit, and the second potential signal is converted into an actual status code of the energy storage.

[0040] The actual status code of the mains power and the actual status code of the energy storage are concatenated bit by bit to generate the current physical connection feature word;

[0041] The current physical connection feature word is filled into the feedback message as valid payload data, and the current topology status feedback is sent.

[0042] Preferably, in the model adaptive correction module, the step of calculating the parameter weight adjustment values ​​of the benchmark running model using an error feedback-based model correction algorithm includes:

[0043] Extract continuous historical numerical deviations within a preset time window from the historical deviation data, and construct a time-series deviation matrix in chronological order.

[0044] The time-series deviation matrix is ​​input into a preset long short-term memory network model, and forward propagation is performed to output the prediction deviation value for the next time step.

[0045] Calculate the squared difference between the predicted deviation value and the numerical deviation currently output by the edge digital twin module, and determine the calculation result as the model prediction loss value;

[0046] The partial derivative of the model's predicted loss value with respect to the current feature weights of the benchmark running model is calculated using the chain rule, and a weight gradient vector is generated.

[0047] The weight gradient vector is multiplied by a preset learning rate scalar, and the result is used as the parameter weight adjustment value.

[0048] Secondly, the present invention provides the following technical solution: a management device combining artificial intelligence with wind, solar and energy storage applications, comprising: a main body, wind power generation components, a generator, an LED display screen, a photovoltaic panel, wires and an integrated controller;

[0049] The wind power generation components are mounted on the top of the main body;

[0050] The generator is connected to the wind power generation component via a transmission connection;

[0051] The photovoltaic panel is disposed on the outer surface of the main body;

[0052] The LED display screen is disposed on the side wall of the main body;

[0053] The integrated controller is located inside the main body;

[0054] The wire connects the generator, the photovoltaic panel, the LED display screen, and the integrated controller.

[0055] The present invention has the following beneficial effects:

[0056] 1. In this invention, by deeply integrating real-time electrical parameters with monitoring video image features, and using weighted Euclidean distance and convolutional neural networks to achieve multi-dimensional state perception, the defects of traditional single threshold monitoring that are easily affected by environmental interference and false alarms are effectively overcome, thereby improving the accuracy of fault identification in wind, solar and energy storage systems.

[0057] 2. In this invention, an adaptive correction module based on a long short-term memory network is used to capture the evolution trend of numerical deviation in real time, and the weights of the benchmark model are automatically fine-tuned through the backpropagation algorithm, thereby realizing online compensation for equipment aging and sensor drift and maintaining the diagnostic performance of the system for long-term operation.

[0058] 3. In this invention, a strategy mapping mechanism based on hash lookup, combined with contact potential feedback technology, accurately converts logical decisions into underlying hardware action instructions and verifies the execution results in real time, forming a closed-loop control from instruction issuance to physical action confirmation, thus ensuring the reliability of the power switching process. Attached Figure Description

[0059] Figure 1 is an architecture diagram of a management system combining artificial intelligence with wind, solar and energy storage applications proposed in this invention;

[0060] Figure 2 is a perspective view of a management system combining artificial intelligence with wind, solar and energy storage applications proposed in this invention.

[0061] The components include: 1. Main body; 2. Wind power generation components; 3. Generator; 4. LED display screen; 5. Photovoltaic panel; 6. Wires; and 7. Integrated controller. Detailed Implementation

[0062] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Example 1:

[0064] In a first embodiment of the present invention, the present invention provides a management system combining artificial intelligence with wind, solar and energy storage applications, as shown in Figure 1, including:

[0065] The multi-source data acquisition module is configured to acquire the electrical operating parameters of the load device, the ambient temperature and humidity data, and the monitoring video frames using their respective independent preset sampling clocks, perform time alignment operations on the electrical operating parameters, the ambient temperature and humidity data, and the monitoring video frames, and output multi-dimensional fusion perception records.

[0066] Furthermore, in the multi-source data acquisition module, the steps of performing time-series alignment operations on electrical operating parameters, ambient temperature and humidity data, and monitoring video frames to output multi-dimensional fused sensing records include:

[0067] The received electrical operating parameters, ambient temperature and humidity data, and monitoring video frames are marked with a timestamp using a preset unified system clock.

[0068] Extract the frame generation timestamp of the surveillance video frame and set the frame generation timestamp as the reference alignment anchor point for data fusion;

[0069] In the time series of electrical operating parameters, two adjacent sampling points located before and after the reference alignment anchor point are retrieved, and the electrical fitting values ​​of the electrical operating parameters at the reference alignment anchor point are calculated using a linear interpolation algorithm.

[0070] In the time series of environmental temperature and humidity data, retrieve sampling points whose absolute value of the time difference with the benchmark alignment anchor point is lower than that of other sampling points, and lock the data of that sampling point as the environmental retention value;

[0071] The image encoding data, electrical fitting values, and environmental maintenance values ​​of the monitoring video frames are encapsulated in a unified protocol format to generate a multi-dimensional fusion sensing record.

[0072] Specifically, the multi-source data acquisition module is deployed on an edge computing gateway device equipped with an industrial-grade ARM architecture processor and an FPGA co-processing unit. After system startup, the local high-precision crystal clock is calibrated via the Network Time Protocol (NTP) to control the system time synchronization error to the microsecond level. The data acquisition process utilizes the FPGA's programmable gate array to process multiple signals in parallel. It connects to a three-phase multi-function power meter via an RS485 isolation interface using the Modbus RTU protocol to read the A-phase voltage, B-phase voltage, C-phase voltage, and total three-phase active power data within the register address range. The sampling frequency is set to 1000Hz. Simultaneously, it reads the ambient temperature and relative humidity values ​​from the embedded temperature and humidity sensor via the I2C bus interface, with a sampling frequency set to 1Hz. It pulls the H.265 encoded video stream from an industrial camera via a gigabit Ethernet interface. The video resolution is 3840x2160 pixels, and the frame rate is 30 frames per second.

[0073] The moment all the raw data packets acquired enter the receive buffer of the Direct Memory Access (DMA) controller, the driver immediately reads the current unified system clock value as the receive timestamp t. recv It is then appended to the packet header, and the data is subsequently pushed into the pre-allocated electrical ring buffer Q in memory. elec , Environmental Circular Buffer Q env and video circular buffer Q video .

[0074] The data fusion engine first starts from the video circular buffer Q video The head of the queue extracts the current video frame data F to be processed. curr The original sampling time in the supplemental enhancement information (SEI) of the video frame is analyzed and defined as the reference alignment anchor point T for data fusion. anchor The engine then locked the electrical annular buffer Q.elec The binary search algorithm is used to quickly locate index k in the time-ordered buffer, such that the electrical sampling time t corresponding to that index is such that... k Electrical sampling time t corresponding to the next index k+1 Satisfying the time interval condition t k ≤T anchor <t k+1 The system calculates the time span Δt = t k+1 -t k It is then compared with a preset validity threshold to prevent interpolation distortion caused by data loss.

[0075] Define an electrical characteristic vector that includes three-phase voltage components and total power components. The system calculates the interpolation weighting coefficient α and the electrical fitting numerical vector E at the reference alignment anchor point time. target The specific formula for vector linear interpolation is as follows:

[0076]

[0077] E target =E k +α·(E k+1 -E k );

[0078] In the formula, t k t is the timestamp of the electrical sampling point at the moment preceding the reference alignment anchor point. k+1 E is the timestamp of the electrical sampling point one moment after the reference alignment anchor point. k For time point t k The collected electrical feature vector, E k+1 For time point t k+1 The collected electrical feature vectors, and the calculation result E target This refers to the electrical fitting values ​​after timing alignment.

[0079] For aligning environmental temperature and humidity data, the system traverses the environmental circular buffer Q. env The effective data window in the dataset is defined as an environmental state vector H = [Temp, Hum] containing temperature and humidity components. T The system constructs a minimum objective function to find the best matching index j that is closest in time to the benchmark alignment anchor point. The objective function formula is as follows:

[0080]

[0081] In the formula, i is the traversal index within the environmental circular buffer, and t iLet T be the timestamp of the i-th environmental sampling point. anchor Using the baseline alignment anchor point, argmin represents the parameter value that minimizes the objective function. The system extracts the corresponding environment state vector H based on the calculated optimal index j. j This value is directly locked to the environmental preservation value.

[0082] Finally, the system allocates a contiguous memory space and sequentially writes the frame header identifier and the reference alignment anchor point T. anchor The calculated electrical fitting numerical vector E target The locked environment preserves the numerical vector H j and video frame F curr The image encoding payload is processed and a checksum is generated using the CRC32 algorithm and appended to the end of the data to complete the encapsulation and output of the multidimensional fusion sensing record.

[0083] By using a vector-based linear interpolation algorithm and a nearest neighbor time search algorithm, the sampling phase deviation between high-frequency electrical data, low-frequency environmental data and mid-frequency video data is eliminated, ensuring strict synchronization of various physical quantities input into the artificial intelligence model in the time dimension and improving the accuracy of subsequent multimodal state analysis.

[0084] The edge digital twin module extracts environmental state features based on multi-dimensional fusion perception records, retrieves theoretical benchmark operation feature vectors corresponding to environmental state features using a preset benchmark operation model, calculates the numerical deviation between electrical operation parameters and theoretical benchmark operation feature vectors, maps the numerical deviations to a preset business state space, and outputs business state classification labels.

[0085] Furthermore, in the edge digital twin module, the steps for calculating the numerical deviation between electrical operating parameters and the theoretical reference operating characteristic vector include:

[0086] Analyze the multi-dimensional fusion sensing records to separate real-time electrical feature vectors and environmental state features;

[0087] By inputting environmental state characteristics into the multidimensional state mapping unit in the benchmark operation model, the theoretical benchmark operation feature vector corresponding to the environmental state characteristics is retrieved.

[0088] For the corresponding dimensions of the real-time electrical feature vector and the theoretical benchmark operating feature vector, the absolute value of the component difference is calculated respectively to generate the difference vector;

[0089] Perform a weighted Euclidean distance operation on the difference vector and determine the scalar result as the numerical deviation.

[0090] Furthermore, in the edge digital twin module, the steps of mapping numerical deviations to a preset business state space and outputting business state classification labels include:

[0091] Extract surveillance video frames from multi-dimensional fusion perception records, input the surveillance video frames into a preset image anomaly recognition neural network, and extract visual anomaly confidence values;

[0092] Construct a joint state feature set that includes numerical bias and visual anomaly confidence values;

[0093] The joint state feature group is input into the pre-trained Softmax classifier to calculate and generate the probability distribution vector corresponding to each sub-state in the preset business state set.

[0094] Extract the state index corresponding to the component whose value is higher than the other components in the probability distribution vector, and convert the state index into a business state classification label in text format.

[0095] Specifically, the edge digital twin module is deployed in the memory of an edge computing controller equipped with a GPU parallel computing unit. The system first initiates a data parsing process, reads the multi-dimensional fusion sensing records output by the previous-level multi-source data acquisition module, performs unpacking operations according to a predefined binary protocol structure, and separates out the real-time electrical feature vector E composed of phase A voltage, phase B voltage, phase C voltage, and total active power. real , and the environmental state characteristics H, which consist of ambient temperature and relative humidity.

[0096] The system calls the pre-loaded benchmark operating model, which is a multidimensional lookup table built based on historical normal operating data. This table covers the theoretical electrical parameters of the load equipment under different temperature and humidity combinations. The system inputs the environmental state feature H into the multidimensional state mapping unit of the benchmark operating model and uses a bilinear interpolation algorithm to retrieve and calculate the corresponding theoretical benchmark operating feature vector E under the current environmental conditions. ref .

[0097] After obtaining the theoretical baseline value, the system targets the real-time electrical characteristic vector E. real Eigenvector E running against theoretical benchmark ref Differential calculations are performed on the corresponding dimensions. The system predefines a weight vector W, where each element corresponds to a weighting coefficient for the influence of voltage and power parameters on the system state. The weight values ​​are determined based on the electrical characteristic specifications of the load equipment. The system calculates the weighted Euclidean distance between the two sets of vectors, using it as the numerical deviation D characterizing the degree of deviation in equipment operation. dev The specific formula for calculating the numerical deviation is as follows:

[0098]

[0099] In the formula, n represents the dimension of the feature vector, and w iLet e ​​be the weight coefficient of the i-th dimension in the weight vector W. real,i Let e ​​be the value of the i-th dimension in the real-time electrical feature vector. ref,i The scalar result D is obtained by operating on the numerical values ​​of the i-th dimension of the eigenvector of the theoretical benchmark. dev This refers to the quantified deviation of electrical operating values.

[0100] The system then extracts image data from the surveillance video frames recorded by multi-dimensional fusion perception, scales them to a size of 224 pixels by 224 pixels, and normalizes the pixel values. This image data is then input into a pre-set image anomaly recognition neural network. This network is constructed as a convolutional neural network structure, sequentially containing a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, and a fully connected layer. The first convolutional layer is configured with 32 convolutional kernels of size 3x3, using the ReLU activation function to extract edge features. The second convolutional layer is configured with 64 kernels of size 3x3. The network uses convolutional kernels to extract texture and shape features. The fully connected layers have 128 nodes, and the output layer contains one neuron connected to a sigmoid activation function. Before deployment, the network was trained offline using anomaly image datasets (including electrical sparks, smoke, and foreign object intrusion) and normal operation image datasets. Iterative parameter updates were performed using a binary cross-entropy loss function and the Adam optimizer until the loss function converged. The system inputs the current frame into the trained network, and the scalar value between 0 and 1 calculated by the output layer is the visual anomaly confidence value C. vis .

[0101] The system constructs a two-dimensional joint state feature set X. joint Its first component is the numerical deviation D obtained from the aforementioned calculation. dev The second component is the confidence score C for visual anomalies. vis The system inputs the joint state feature group into a pre-trained Softmax classifier, which is constructed as a three-layer feedforward neural network. The input layer contains 2 nodes, the hidden layer contains 10 nodes using the ReLU activation function, and the output layer contains 4 nodes, corresponding to four service states: normal power supply, overload warning, environmental anomaly, and short circuit fault. The classifier is pre-trained using a large amount of labeled historical joint feature data and uses multi-class cross-entropy as the loss function. The system performs forward propagation through the network and uses the Softmax function in the output layer to calculate and generate the probability distribution vector P, as shown in the following formula:

[0102]

[0103] In the formula, P k z represents the probability that the current system is in the k-th business state. kThe output layer's k-th node is the weighted sum of inputs, where K is the total number of preset business state sets (4). The system iterates through and calculates the probability distribution vector P, extracting the index key k corresponding to the component with the largest value. max The system accesses the pre-stored state index mapping table and sets the index key value k. max Convert the data into the corresponding text format business status classification label, and use this label as the final output of the edge digital twin module.

[0104] By quantifying electrical parameter drift using weighted Euclidean distance and capturing non-electrical features using a visual network, high-precision identification of the operating status of load equipment under complex working conditions is achieved.

[0105] The power supply strategy mapping module, based on the service status classification label, retrieves the target power supply mode corresponding to the service status classification label from the pre-stored status-action mapping table, and generates a dual-path switching combination instruction that includes the target conduction status of the mains circuit breaker and the target conduction status of the energy storage circuit breaker.

[0106] Furthermore, in the power supply strategy mapping module, the step of generating a dual-path switching combination command that includes the target conduction state of the mains circuit breaker and the target conduction state of the energy storage circuit breaker includes:

[0107] Extract the index key value of the business status category label;

[0108] Using the index key as the search keyword, locate the target control strategy row in the status-action mapping table, and read the preset mains circuit control identifier and energy storage circuit control identifier in the target control strategy row;

[0109] Construct an instruction data packet containing a frame header sequence, device physical address, and function code;

[0110] Write the mains circuit control identifier into the mains control bit of the instruction data packet, and write the energy storage circuit control identifier into the energy storage control bit of the instruction data packet to generate a dual-path switching combination instruction.

[0111] Specifically, the power supply strategy mapping module runs within the real-time logic tasks of the edge controller. The system pre-loads a structured state-action mapping table T into the EEPROM non-volatile memory. map The mapping table is constructed as a hash graph data structure, where each row of data defines a rigid correspondence between a specific input state and an output action. The specific structural form is described as {Index:[Flag grid Flag store ]}.

[0112] To ensure the completeness of the control logic, this mapping table predefines specific mapping instances: for example, when the key value Index is an integer 0, it corresponds to "normal mains power direct supply mode," and the mains circuit control flag Flag in the value field is... grid The energy storage loop control flag is preset to a value of 1. store The default value is 0; when the key value Index is an integer 1, it corresponds to "Off-grid mode in case of mains power failure", and the Flag value in this field is... grid Flag is preset to the value 0. store The default value is 1; when the key value Index is an integer 2, it corresponds to "critical short circuit protection mode", and the Flag value field will be set accordingly. grid With Flag store All values ​​are forcibly preset to 0 in order to achieve dual-path cutoff.

[0113] The module receives in real time the index key value k corresponding to the service status classification label output by the previous-level edge digital twin module. in Using a hash lookup algorithm in the state-action mapping table T map Quickly locate the key value equal to k in The target control strategy line is used to read the specific mains circuit control identifier u. grid and energy storage circuit control marking u store The values ​​of these two identifiers are strictly limited to binary Boolean values, where the value 1 represents a closed conducting state and the value 0 represents an open cutoff state.

[0114] The system then initializes an 8-byte instruction data packet buffer conforming to the Modbus RTU transmission standard. The system sequentially writes the hexadecimal value 0x01 as the device physical address, 0x0F as the write multi-coil function code, and 0x0000 as the start register address to the buffer header. To integrate the separate control flags into a hardware-recognizable instruction format, the system performs a bit-weighted summation operation on the read control flags to generate the target control word B. ctrl The specific bit mapping operation formula is as follows:

[0115] B ctrl =u grid ·2 0 +u store ·2 1 ;

[0116] In the formula u grid For the mains circuit control identifier read from the table, u store For the energy storage loop control identifier read from the table, 2 0 This represents the bit weight of Bit0, the least significant bit in the data byte, corresponding to the mains relay drive bit, 2.1 The bit weight representing the second least significant bit Bit1 corresponds to the energy storage relay drive bit, and the generated B is calculated. ctrl This is a compressed encoded byte containing the target state of the dual circuit breakers. The system writes this control word into the data payload segment of the instruction data packet, calls the CRC16 algorithm to calculate the check code of the entire data packet and appends it to the end, generates the final dual-circuit switching combination instruction and pushes it into the RS485 transmission queue.

[0117] By pre-setting a lookup mechanism containing specific state instances in non-volatile memory, combined with deterministic bit-level encoding operations, the abstract business state decisions of the upper layer are accurately converted into binary control instructions executable by the underlying hardware, ensuring that the switching logic of the mains power and energy storage circuits under different operating conditions is rigorous and traceable.

[0118] The physical link execution module, based on dual-path switching combination commands, drives the mains circuit breaker and the energy storage circuit breaker to operate to the target conduction state defined by the commands, collects the contact signals after the circuit breaker operates, and outputs the current topology status feedback.

[0119] Furthermore, in the physical link execution module, the steps for driving the mains circuit breaker and the energy storage circuit breaker to operate to the target conduction state specified by the instruction include:

[0120] Analyze the dual-channel switching combination command and separate the mains control bit and energy storage control bit contained within it;

[0121] The separated mains control bits are converted into mains drive voltage signals, and the separated energy storage control bits are converted into energy storage drive voltage signals.

[0122] The mains drive voltage signal is applied to the operating coil of the mains circuit breaker through the first isolation drive channel, driving the moving contact of the mains circuit breaker to move to the mains target position corresponding to the target conduction state of the mains circuit breaker.

[0123] Simultaneously, the energy storage drive voltage signal is applied to the operating coil of the energy storage circuit breaker through the second isolation drive channel, driving the moving contact of the energy storage circuit breaker to move to the energy storage target position corresponding to the target conduction state of the energy storage circuit breaker.

[0124] Furthermore, in the physical link execution module, the steps for outputting the current topology status feedback include:

[0125] Collect the first potential signal of the auxiliary contacts of the mains circuit breaker and the second potential signal of the auxiliary contacts of the energy storage circuit breaker;

[0126] The first potential signal is converted into the actual status code of the mains power using a preset signal acquisition circuit, and the second potential signal is converted into the actual status code of the energy storage.

[0127] The actual status code of the mains power and the actual status code of the energy storage are concatenated bit by bit to generate the current physical connection feature word;

[0128] The current physical connection characteristic is filled into the feedback message as valid payload data, and the current topology status feedback is sent.

[0129] Specifically, the physical link execution module is integrated into the core control board of the power control cabinet. This board includes a microcontroller unit (MCU), an opto-isolated drive circuit, and a status feedback acquisition circuit. The system first receives dual-channel switching combination instruction data packets from the power supply strategy mapping module via an RS485 communication bus. After the MCU verifies the integrity of the data packet using a cyclic redundancy check algorithm, it parses the data packet payload segment, extracts the control byte containing the target conduction state, and performs bitmasking operations to separate the least significant bit of the control byte as the mains control bit u. grid_cmd The second least significant bit is separated as the energy storage control bit u. store_cmd The MCU's general purpose input / output (GPIO) ports output corresponding logic level signals based on these two binary control bits. When the control bit is 1, it outputs a 3.3-volt high level, and when it is 0, it outputs a 0-volt low level.

[0130] The logic level signal output by the MCU is transmitted to the power drive stage via a high-speed optocoupler isolator. The optocoupler isolator realizes electrical isolation between the low-voltage control side and the high-voltage drive side. For the mains circuit, when the optocoupler is turned on, it triggers the subsequent N-channel MOSFET power transistor to saturate and turn on, applying the onboard 24-volt DC power supply to both ends of the electromagnetic operating coil of the mains circuit breaker. After the coil is energized, it generates electromagnetic attraction, driving the moving contact of the circuit breaker to overcome the resistance of the return spring and move to the closed position where it contacts the stationary contact, thereby realizing the physical conduction of the mains circuit. For the energy storage circuit, the system converts the level signal corresponding to the energy storage control position into a drive voltage through a second isolation drive channel with the same structure, independently controlling the opening and closing actions of the energy storage circuit breaker.

[0131] After the driving action is completed, the system immediately initiates the status feedback acquisition process. The physical link execution module is connected to the auxiliary contacts of the mains circuit breaker and the energy storage circuit breaker via a dedicated signal line. The auxiliary contacts maintain synchronous action with the main contacts through a mechanical linkage. When the main contacts are closed, the auxiliary contacts are connected, introducing the external detection voltage into the acquisition circuit; when the main contacts are open, the auxiliary contacts are disconnected, and the input terminal of the acquisition circuit is clamped to ground potential by a pull-down resistor. The MCU's built-in analog-to-digital converter (ADC) acquires the first potential signal V of the mains auxiliary contact circuit. sense1 and the second potential signal V of the energy storage auxiliary contact circuit sense2The system uses voltage comparison logic to convert analog voltage signals into digital status codes. The specific conversion calculation formula is as follows:

[0132]

[0133] In the formula, S grid This is the converted AC power status code, where 1 represents closed and 0 represents open; V sense1 The collected mains auxiliary contact voltage value; V thresh The preset logic decision threshold voltage is set to 12 volts, and the actual energy storage status code is S. store The conversion process is the same, using the second potential signal V. sense2 A judgment will be rendered.

[0134] The system then concatenates the two independent status codes at the binary bit level to generate the current physical connection feature word W, which represents the current physical topology of the system. status The calculation formula is as follows:

[0135] W status =(S grid ·2 0 )+(S store ·2 1 );

[0136] 2 in the formula 0 The least significant bit of the binary number is the bit weight coefficient, 2. 1 The weight coefficient of the second least significant bit in binary is used to calculate W. status It is a decimal integer, and its lower two bits in binary form precisely correspond to the actual on / off state of the mains power and energy storage circuit breaker. The MCU encapsulates this feature word into the payload area of ​​the feedback message and sends it to the upper management system through the communication interface.

[0137] The opto-isolation mechanism enables safe operation of the high-voltage circuit breaker. By utilizing the voltage feedback of the auxiliary contacts and digital coding logic, a closed-loop physical state confirmation mechanism is formed, ensuring that the execution results of control commands can be perceived by the system in real time and accurately.

[0138] The model adaptive correction module, based on the current topology feedback and historical deviation data formed by the accumulation of numerical deviations, calculates the parameter weight adjustment values ​​of the benchmark running model through the error feedback-based model correction algorithm, and writes the parameter weight adjustment values ​​as model update coefficients into the edge digital twin module.

[0139] Furthermore, in the model adaptive correction module, the steps for calculating the parameter weight adjustment values ​​of the baseline running model using an error feedback-based model correction algorithm include:

[0140] Extract continuous historical numerical deviations within a preset time window from historical deviation data, and construct a time-series deviation matrix in chronological order.

[0141] The timing bias matrix is ​​input into a pre-set long short-term memory network model, forward propagation is performed, and the prediction bias value for the next time step is output.

[0142] Calculate the squared difference between the prediction deviation value and the numerical deviation of the current output of the edge digital twin module, and determine the calculation result as the model prediction loss value;

[0143] The chain rule is used to calculate the partial derivative of the model's predicted loss value with respect to the current feature weights of the benchmark model, and then the weight gradient vector is generated.

[0144] Perform a multiplication operation between the weight gradient vector and the preset learning rate scalar, and determine the result as the parameter weight adjustment value.

[0145] Specifically, the model adaptive correction module runs as a background daemon on the neural processing unit (NPU) of the edge computing gateway. This module allocates a 50-level first-in-first-out circular queue in memory specifically for storing the numerical deviation data output by the edge digital twin module at continuous sampling times. The system monitors the output port of the edge digital twin module in real time, and whenever a new numerical deviation d is generated... curr The system will push it to the tail of the circular queue, while removing the oldest data from the front of the queue to maintain a constant queue length.

[0146] The system extracts the numerical deviation data of the most recent 50 time steps from the circular queue and constructs a one-dimensional time series deviation matrix D with a dimension of 50×1 according to the chronological order. seq .

[0147] The system calls a Long Short-Term Memory (LSTM) network model pre-loaded in the NPU memory. This model is a recurrent neural network built on the TensorFlow framework. Its specific structure includes an input layer, a hidden layer, and an output layer. The input layer has 1 node, corresponding to a single-variable time series input. The hidden layer contains 128 LSTM neurons, each of which integrates a forget gate, an input gate, and an output gate. Tanh is used as the state activation function, and Sigmoid is used as the gating activation function to capture the long-term evolution trend of the biased data. The output layer contains one fully connected node and uses a linear activation function.

[0148] Before system deployment, the model was trained offline using numerical deviation sequences extracted from historical operation logs as a sample set. During training, mean squared error (MSE) was used as the loss function, and the Adam optimizer was used to iterate the parameters at a learning rate of 0.001 until the validation set loss value no longer decreased. This enabled the model to predict the theoretical deviation at the next moment based on historical trends.

[0149] The system will construct the timing deviation matrix D seq The input is fed into the trained LSTM model, the forward propagation operation is performed, and the prediction bias value for the next time step is output. At this moment, the system reads the actual numerical deviation d of the edge digital twin module at the current moment. curr The squared difference between the predicted deviation and the actual deviation is calculated and defined as the model prediction loss value L. The calculation formula is as follows:

[0150]

[0151] in the formula d represents the theoretical bias in the predictions made by the LSTM network based on historical trends. curr The coefficient represents the actual deviation calculated based on the physical model at the current moment. This simplifies subsequent differentiation operations.

[0152] To correct the incompatibility of the feature weight vector W in the baseline operating model with current equipment aging or environmental drift, the system uses the chain rule to calculate the model prediction loss value L for each component w in the weight vector. i The partial derivatives, and a review of the numerical bias d in the edge digital twin module. curr The computational logic, namely the weighted Euclidean distance algorithm, derives the following gradient calculation formula:

[0153]

[0154] in the formula Let e ​​be the gradient value of the i-th component of the feature weight vector. real,i Let e ​​be the value of the i-th dimension of the real-time electrical feature vector in the edge digital twin module. ref,i The value of the i-th dimension of the eigenvector is used as the theoretical benchmark.

[0155] The system generates a weight gradient vector based on the calculated gradient values. Finally, the system defines a positive real number as the learning rate scalar η, with a value of 0.01. The weight gradient vector is multiplied by the learning rate scalar, and the result is reversed to minimize the error, yielding the parameter weight adjustment vector ΔW. The calculation formula is as follows:

[0156]

[0157] The system writes the calculated parameter weight adjustment value ΔW into the configuration register of the edge digital twin module, and in the next calculation cycle, it adds the adjustment value to the current feature weight vector W to complete the online update of the weights.

[0158] The gradient descent algorithm enables online adaptive fine-tuning of physical model parameters, effectively compensating for the decrease in model accuracy caused by equipment aging or sensor drift, and ensuring the robustness of the system in long-term operation.

[0159] Example 2:

[0160] As shown in Figure 2, the present invention also provides a management device that combines artificial intelligence with wind, solar and energy storage applications, including: a main body 1, a wind power generation component 2, a generator 3, an LED display screen 4, a photovoltaic panel 5, a wire 6 and an integrated controller 7;

[0161] The wind power generation component 2 is installed on top of the main body 1;

[0162] Generator 3 is connected to wind power generation component 2 via a transmission.

[0163] Photovoltaic panel 5 is disposed on the outer surface of main body 1;

[0164] LED display screen 4 is installed on the side wall of the main body 1;

[0165] The integrated controller 7 is located inside the main body 1;

[0166] The wire 6 connects the generator 3, the photovoltaic panel 5, the LED display screen 4, and the integrated controller 7.

[0167] Specifically, after the device is started, the wind power generation component 2 captures natural wind energy and drives the generator 3 to generate alternating current. At the same time, the photovoltaic panel 5 receives sunlight and generates direct current. The integrated controller 7 uses a preset unified system clock to sample the electrical operating parameters such as voltage and current output by the generator 3 and the photovoltaic panel 5 in real time, and simultaneously acquires the current ambient temperature and humidity data and monitoring video frames. Subsequently, the processor inside the integrated controller 7 performs time alignment operation on the data from the above different sources, and generates a standardized multi-dimensional fusion perception record through linear interpolation and timestamp anchoring, completing the conversion of the physical world into digital signals.

[0168] Based on this, the integrated controller 7 calls the internally preset benchmark operation model, inputs the collected environmental state features into the model, retrieves the theoretical benchmark operation feature vector that should be achieved under the current environment, and then calculates the numerical deviation between the actual collected electrical operation parameters and the theoretical benchmark, and maps the deviation to the preset business state space. Combined with the image anomaly recognition results, the integrated controller 7 finally outputs a business state classification label for the current moment.

[0169] Based on the determined business status classification labels, the integrated controller 7 consults the internally stored state-action mapping table to generate a dual-path switching combination command containing the on / off logic of the mains power and energy storage circuits. The controller drives the internal power circuit to operate according to the command, automatically adjusting the flow of power between the generator 3 and the photovoltaic panel 5, and controlling the on / off of the mains power circuit. At the same time, the integrated controller 7 sends the current business status classification labels and topology status to the LED display screen 4 through the wire 6, so that the LED display screen 4 displays the operating status of the device in real time. In addition, during continuous operation, the integrated controller 7 records the accumulated historical deviation data, uses the long short-term memory network to reverse calculate the predicted loss value, and automatically updates the parameter weights of the baseline operating model, thereby realizing the adaptive iteration of the model.

[0170] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A management system combining artificial intelligence with wind, solar, and energy storage applications, characterized in that, include: The multi-source data acquisition module is configured to acquire electrical operating parameters of the load device, ambient temperature and humidity data, and monitoring video frames using their respective independent preset sampling clocks, perform time alignment operations on the electrical operating parameters, ambient temperature and humidity data, and monitoring video frames, and output multi-dimensional fusion perception records. The edge digital twin module extracts environmental state features based on the multi-dimensional fusion perception record, retrieves the theoretical benchmark operation feature vector corresponding to the environmental state features using a preset benchmark operation model, calculates the numerical deviation between the electrical operation parameters and the theoretical benchmark operation feature vector, maps the numerical deviation to a preset business state space, and outputs business state classification labels. The power supply strategy mapping module, based on the service status classification label, retrieves the target power supply mode corresponding to the service status classification label from the pre-stored status-action mapping table, and generates a dual-path switching combination command including the target conduction state of the mains circuit breaker and the target conduction state of the energy storage circuit breaker; the physical link execution module, based on the dual-path switching combination command, drives the mains circuit breaker and the energy storage circuit breaker to act to the target conduction state defined by the command, collects the contact signals after the circuit breaker action, and outputs the current topology status feedback; the model adaptive correction module, based on the current topology status feedback and the historical deviation data formed by the accumulation of numerical deviations, calculates the parameter weight adjustment value of the benchmark operating model through an error feedback-based model correction algorithm, and writes the parameter weight adjustment value as the model update coefficient into the edge digital twin module.

2. The management system combining artificial intelligence with wind, solar, and energy storage applications according to claim 1, characterized in that, In the multi-source data acquisition module, the step of performing time-series alignment operations on the electrical operating parameters, the ambient temperature and humidity data, and the monitoring video frames to output a multi-dimensional fusion sensing record includes: marking the received electrical operating parameters, ambient temperature and humidity data, and monitoring video frames with a preset unified system clock; extracting the frame generation timestamp of the monitoring video frame and setting the frame generation timestamp as the reference alignment anchor point for data fusion; retrieving two adjacent sampling points before and after the reference alignment anchor point in the time series of the electrical operating parameters, and calculating the electrical fitting value of the electrical operating parameters at the reference alignment anchor point using a linear interpolation algorithm; retrieving sampling points in the time series of the ambient temperature and humidity data whose absolute value of the time difference with the reference alignment anchor point is lower than that of other sampling points, and locking the data of the sampling point as the environmental retention value; and encapsulating the image encoding data of the monitoring video frame, the electrical fitting value, and the environmental retention value according to a unified protocol format to generate the multi-dimensional fusion sensing record.

3. The management system combining artificial intelligence with wind, solar, and energy storage applications according to claim 1, characterized in that, In the edge digital twin module, the step of calculating the numerical deviation between the electrical operating parameters and the theoretical benchmark operating feature vector includes: parsing the multi-dimensional fused sensing record to separate the real-time electrical feature vector and environmental state features; inputting the environmental state features into the multi-dimensional state mapping unit in the benchmark operating model to retrieve the theoretical benchmark operating feature vector corresponding to the environmental state features; calculating the absolute value of the component difference for the corresponding dimensions of the real-time electrical feature vector and the theoretical benchmark operating feature vector to generate a difference vector; performing a weighted Euclidean distance operation on the difference vector, and determining the scalar result obtained from the operation as the numerical deviation.

4. The management system combining artificial intelligence with wind, solar, and energy storage applications according to claim 1, characterized in that, In the edge digital twin module, the step of mapping the numerical deviation to a preset business state space and outputting a business state classification label includes: extracting monitoring video frames from the multi-dimensional fusion perception record; inputting the monitoring video frames into a preset image anomaly recognition neural network to extract visual anomaly confidence values; constructing a joint state feature group containing the numerical deviation and the visual anomaly confidence values; inputting the joint state feature group into a pre-trained Softmax classifier to calculate and generate probability distribution vectors corresponding to each sub-state in the preset business state set; extracting the state index corresponding to the component whose value is higher than the other components in the probability distribution vector; and converting the state index into a text-formatted business state classification label.

5. The management system combining artificial intelligence with wind, solar, and energy storage applications according to claim 1, characterized in that, In the power supply strategy mapping module, the step of generating a dual-path switching combination instruction containing the target on-state of the mains circuit breaker and the target on-state of the energy storage circuit breaker includes: extracting the index key value of the service status classification label; using the index key value as the search keyword, locating the target control strategy row in the status-action mapping table, and reading the preset mains circuit control identifier and energy storage circuit control identifier in the target control strategy row; constructing an instruction data packet containing a frame header sequence, device physical address, and function code; writing the mains circuit control identifier into the mains control bit of the instruction data packet, writing the energy storage circuit control identifier into the energy storage control bit of the instruction data packet, and generating a dual-path switching combination instruction.

6. The management system combining artificial intelligence with wind, solar, and energy storage applications according to claim 1, characterized in that, In the physical link execution module, the steps of driving the mains circuit breaker and the energy storage circuit breaker to operate to the target conduction state defined by the instruction include: parsing the dual-path switching combination instruction and separating the mains control bit and the energy storage control bit contained therein; converting the separated mains control bit into a mains drive voltage signal and the separated energy storage control bit into an energy storage drive voltage signal; loading the mains drive voltage signal onto the operating coil of the mains circuit breaker through a first isolation drive channel, driving the moving contact of the mains circuit breaker to move to the mains target position corresponding to the target conduction state of the mains circuit breaker; and simultaneously loading the energy storage drive voltage signal onto the operating coil of the energy storage circuit breaker through a second isolation drive channel, driving the moving contact of the energy storage circuit breaker to move to the energy storage target position corresponding to the target conduction state of the energy storage circuit breaker.

7. The management system combining artificial intelligence with wind, solar, and energy storage applications according to claim 1, characterized in that, In the physical link execution module, the step of outputting the current topology status feedback includes: acquiring the first potential signal of the auxiliary contact of the mains circuit breaker and the second potential signal of the auxiliary contact of the energy storage circuit breaker; using a preset signal acquisition circuit to convert the first potential signal into a mains actual status code and the second potential signal into an energy storage actual status code; concatenating the mains actual status code and the energy storage actual status code bit by bit to generate the current physical connection feature word; filling the current physical connection feature word as valid load data into the feedback message and sending the current topology status feedback.

8. The management system combining artificial intelligence with wind, solar, and energy storage applications according to claim 1, characterized in that, In the model adaptive correction module, the step of calculating the parameter weight adjustment value of the benchmark running model using an error feedback-based model correction algorithm includes: extracting continuous historical numerical deviations within a preset time window from the historical deviation data, and constructing a temporal deviation matrix in chronological order; inputting the temporal deviation matrix into a preset long short-term memory network model, performing forward propagation, and outputting the prediction deviation value for the next time step; calculating the squared difference between the prediction deviation value and the numerical deviation currently output by the edge digital twin module, and determining the calculation result as the model prediction loss value; calculating the partial derivative of the model prediction loss value with respect to the current feature weights of the benchmark running model using the chain rule, and generating a weight gradient vector; performing a multiplication operation between the weight gradient vector and a preset learning rate scalar, and determining the operation result as the parameter weight adjustment value.

9. A management device combining artificial intelligence with wind, solar, and energy storage applications, characterized in that, A management system for combining artificial intelligence with wind, solar and energy storage applications as described in any one of claims 1-8, comprising: a main body (1), a wind power generation component (2), a generator (3), an LED display screen (4), a photovoltaic panel (5), wires (6), and an integrated controller (7); the wind power generation component (2) is disposed on the top of the main body (1); the generator (3) is connected to the wind power generation component (2); the photovoltaic panel (5) is disposed on the outer surface of the main body (1); the LED display screen (4) is disposed on the side wall of the main body (1); the integrated controller (7) is disposed inside the main body (1); the wires (6) connect the generator (3), the photovoltaic panel (5), the LED display screen (4), and the integrated controller (7).

Images