An intelligent voltage monitoring system with self-adaptive calibration function

Abstract

The application provides a kind of intelligent voltage monitoring system with adaptive calibration function, it is related to data processing technical field, the system includes: voltage acquisition module;Environment monitoring module;Intelligent drift modeling module is used to train associated data based on neural network model, learn nonlinear drift law under the condition of multi-factor coupling, and construct drift prediction model;Prediction calibration module is used to generate calibration vector for uncollected original voltage data, and dynamically superimposes calibration in real-time sampling process, generates prediction calibration voltage data;Closed loop self-learning module is used to extract deviation trend based on the periodic deviation of prediction calibration voltage and reference voltage, and accordingly online updates drift prediction model;Application output module is used to provide prediction calibration voltage data to the control unit of power equipment;The application improves the autonomy and accuracy of intelligent voltage monitoring.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an intelligent voltage monitoring system with adaptive calibration function. Background Technology

[0002] Existing voltage monitoring systems have incorporated various compensation methods to reduce the impact of sensor parameter drift due to environmental factors. For example, one method utilizes periodic manual calibration, where maintenance personnel recalibrate the sensors using a standard power signal within specific maintenance cycles. Another method introduces a temperature compensation module into the circuit, correcting voltage readings by adding a thermistor to the voltage measurement link or using a software temperature compensation curve. Furthermore, some intelligent monitoring devices attempt to utilize dual-channel redundant sampling, determining the degree of drift by comparing the differences between two independent sensors. These measures have improved the accuracy of voltage monitoring to some extent.

[0003] However, these compensation schemes still have shortcomings in scenarios involving long-term operation of photovoltaic inverters. Manual calibration requires shutdown, making it unsuitable for unattended distributed photovoltaic power plants. While temperature compensation modules can correct for single temperature factors, they struggle to maintain accuracy under multi-factor coupling (such as the combined effects of temperature and aging). Furthermore, dual-channel redundant sampling significantly increases hardware costs and power consumption in outdoor inverters. For example, after 18 months of continuous operation, some inverters, despite temperature compensation, still experienced voltage monitoring errors that gradually increased from ±0.5% to over ±2%, leading to distortion of the voltage reference obtained by the inverter's maximum power point tracking (MPPT) algorithm and directly impacting power generation efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent voltage monitoring system with adaptive calibration function, which aims to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] An intelligent voltage monitoring system with adaptive calibration function, the system comprising:

[0007] The voltage acquisition module is used to receive real-time voltage signals from the monitored power equipment, perform analog-to-digital conversion, and generate raw voltage data.

[0008] The environmental monitoring module is used to synchronously collect operating environmental parameters and store them in time correspondence with the raw voltage data to form environmental-voltage correlation data.

[0009] The intelligent drift modeling module is used to train environmental-voltage correlation data based on a neural network model, learn the nonlinear drift law under multi-factor coupling conditions, construct a drift prediction model, and generate prediction model parameters representing voltage drift trends.

[0010] The prediction calibration module is used to generate a calibration vector based on the prediction model parameters and adaptively adjust the calibration amplitude according to the changing characteristics of the voltage signal; and calculate the calibration confidence level based on the residual between the calibration result and the model predicted voltage value. When the confidence level is lower than the preset confidence threshold, a retraining trigger signal is sent to the closed-loop self-learning module to generate prediction calibration voltage data.

[0011] The closed-loop self-learning module is used to extract the deviation trend based on the periodic deviation between the predicted calibration voltage and the reference voltage, and update the drift prediction model online accordingly. The closed-loop self-learning module also includes a forgetting suppression mechanism to maintain long-term effective knowledge, and triggers incremental learning to achieve model evolution and update when the deviation trend continues to exceed the limit.

[0012] The application output module is used to provide the predicted calibration voltage data to the control unit of the power equipment to support intelligent decision-making for photovoltaic inverter maximum power point tracking control, electric vehicle charging power regulation, or grid safety protection.

[0013] The above-described solution of the present invention has at least the following beneficial effects:

[0014] 1. Achieve automatic voltage calibration without downtime, avoiding the maintenance costs and intermittent risks of traditional manual calibration: Existing methods rely on periodic manual calibration, which requires downtime during maintenance cycles and cannot adapt to the unattended scenarios of distributed photovoltaic power stations; The system of this invention acquires operating data in real time through voltage acquisition and environmental monitoring modules, and the intelligent drift modeling module learns the voltage deviation pattern; the predictive calibration module performs online calibration based on predictive model parameters, thereby achieving automatic voltage compensation during continuous equipment operation; Therefore, it can significantly reduce manual intervention and eliminate power generation losses caused by downtime calibration.

[0015] 2. It can accurately model voltage drift under multi-factor coupling environments, overcoming the single-factor limitation of traditional temperature compensation methods: Traditional circuit temperature compensation can only handle a single temperature factor, while under long-term operating conditions, voltage drift is usually caused by multiple factors such as temperature, humidity, aging, and heat dissipation efficiency. This invention constructs complete environment-voltage correlation data through an environmental monitoring module, and uses a neural network to learn the nonlinear coupling relationship of multiple factors in the intelligent drift modeling module, realizing active learning of complex drift mechanisms, thereby significantly improving drift prediction accuracy. This solution can avoid the failure of traditional methods in high-temperature aging scenarios and ensure the stability of monitoring accuracy in long-term operation.

[0016] 3. Significantly improves voltage measurement reliability in rapidly fluctuating scenarios through adaptive calibration mechanism: Existing compensation techniques are difficult to maintain stable performance in the face of rapid voltage fluctuations (such as when the inverter is dynamically tracking MPPT); This invention introduces a strategy in the predictive calibration module to dynamically adjust the calibration amplitude according to the voltage change characteristics, so that the calibration process can adapt to different fluctuation ranges and achieve timely compensation for rapid disturbances; This scheme avoids the hysteresis and overcompensation problems of traditional static compensation methods in fluctuating scenarios, and improves the real-time performance and tracking capability of voltage calibration.

[0017] 4. Improving long-term model reliability through residual confidence level monitoring for self-diagnosis and automatic retraining: Traditional monitoring equipment cannot determine whether a compensation model has degraded. This invention calculates the residual between the calibration result and the predicted value and generates a calibration confidence level. When the confidence level falls below a threshold, retraining is automatically triggered. This mechanism enables self-identification and self-repair of the model state, maintaining voltage monitoring accuracy over a long period and preventing the gradual accumulation of monitoring errors due to model degradation.

[0018] 5. By using a bias trend-driven online update and forgetting suppression mechanism, the model can evolve over a long period without catastrophic forgetting: The background technology does not include the model's continuous learning ability, which leads to a gradual decline in the accuracy of traditional compensation methods over long-term operation; the closed-loop self-learning module of this invention extracts the periodic bias trend and performs online updates to the prediction model based on the trend; at the same time, a forgetting suppression mechanism is used to distinguish between stable weights and sensitive weights, effectively avoiding catastrophic forgetting that is common in incremental learning; thus, the sustainable evolution and long-term stability maintenance of the model are achieved.

[0019] 6. It can trigger incremental learning when the deviation accumulates abnormally, so as to achieve rapid adaptation to extreme operating conditions: When outdoor inverters experience extreme high temperature, strong radiation or sudden changes in component aging, the voltage drift pattern may change suddenly; traditional compensation methods have limited ability to cope with this and usually require manual intervention; this invention automatically triggers the incremental learning mechanism when the deviation trend continues to exceed the limit, so that the prediction model can complete parameter reconstruction without interrupting operation, quickly adapt to new environmental conditions, and greatly improve the robustness of the system.

[0020] 7. The closed-loop data throughout the entire process makes the monitoring results more compatible with downstream control strategies, improving the energy efficiency of photovoltaic inverters and charging equipment: Traditional compensation error accumulation will directly cause the inverter MPPT curve to shift, resulting in power generation loss; This invention drives the application output module by predicting and calibrating voltage data, and provides the voltage results after multi-level calibration to the MPPT control unit, electric vehicle charging control unit or grid protection module, so that the control strategy operates based on more accurate voltage information; ultimately, it can reduce maximum power point offset, improve power generation efficiency, and enhance the safety of electric vehicle charging and grid operation.

[0021] 8. The overall system can achieve high-precision voltage monitoring on a low-cost hardware basis, avoiding the cost and power consumption increase caused by redundant sensors: Although dual-channel redundant sampling can improve accuracy, it will significantly increase hardware costs; This invention improves accuracy through intelligent modeling, adaptive calibration and self-learning mechanisms, fundamentally replacing the redundant sampling structure, so that the system can achieve near-multi-channel stability and accuracy on a single-channel sampling link; Therefore, it reduces hardware costs, size and power consumption, and is particularly suitable for distributed photovoltaic inverters deployed in batches. Attached Figure Description

[0022] Figure 1 This is an architecture diagram of an intelligent voltage monitoring system with adaptive calibration function provided by an embodiment of the present invention. Detailed Implementation

[0023] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0024] like Figure 1 As shown, an embodiment of the present invention proposes an intelligent voltage monitoring system with adaptive calibration function, the system comprising:

[0025] The voltage acquisition module is used to receive real-time voltage signals from the monitored power equipment, perform analog-to-digital conversion, and generate raw voltage data.

[0026] The environmental monitoring module is used to synchronously collect operating environmental parameters and store them in time correspondence with the raw voltage data to form environmental-voltage correlation data.

[0027] The intelligent drift modeling module is used to train environmental-voltage correlation data based on a neural network model, learn the nonlinear drift law under multi-factor coupling conditions, construct a drift prediction model, and generate prediction model parameters representing voltage drift trends.

[0028] The prediction calibration module is used to generate a calibration vector based on the prediction model parameters and adaptively adjust the calibration amplitude according to the changing characteristics of the voltage signal; and calculate the calibration confidence level based on the residual between the calibration result and the model predicted voltage value. When the confidence level is lower than the preset confidence threshold, a retraining trigger signal is sent to the closed-loop self-learning module to generate prediction calibration voltage data.

[0029] The closed-loop self-learning module is used to extract the deviation trend based on the periodic deviation between the predicted calibration voltage and the reference voltage, and update the drift prediction model online accordingly. The closed-loop self-learning module also includes a forgetting suppression mechanism to maintain long-term effective knowledge, and triggers incremental learning to achieve model evolution and update when the deviation trend continues to exceed the limit.

[0030] The application output module is used to provide the predicted calibration voltage data to the control unit of the power equipment to support intelligent decision-making for photovoltaic inverter maximum power point tracking control, electric vehicle charging power regulation, or grid safety protection.

[0031] In this embodiment of the invention, through the overall design of voltage acquisition, environmental monitoring, intelligent modeling, adaptive calibration, and closed-loop self-learning, the system can continuously obtain accurate and stable voltage monitoring results under long-term operating conditions, avoiding the problem of traditional compensation methods failing easily in unattended and multi-factor coupled environments. The voltage acquisition process, while ensuring signal isolation and high-precision sampling, forms basic data that can be used for model training. The environmental monitoring process, by synchronously acquiring multi-dimensional environmental parameters and aligning them with the voltage data in time, enables a complete characterization of the multiple sources of voltage drift. The drift modeling process utilizes neural networks to learn the nonlinear relationship between environmental factors and voltage drift, thereby accurately predicting drift trends even under complex operating conditions, providing a precise basis for subsequent calibration.

[0032] During calibration, the system adaptively adjusts the calibration amplitude based on the changing characteristics of the voltage signal, enabling the compensation process to be applicable to both rapidly fluctuating and slowly changing operating scenarios. This effectively reduces the undercompensation or overcompensation phenomena that are prone to occur in static compensation methods. By analyzing the residuals between the predicted calibration results and the model output, the system can generate a confidence level reflecting the current calibration reliability. This allows the model to automatically determine whether it is still within its effective operating range. When the calibration results become unreliable, the system can promptly initiate a model update mechanism to avoid the continuous accumulation of monitoring errors due to model degradation, thereby ensuring the long-term reliability of the calibration results.

[0033] During operation, by analyzing the periodic deviation between the predicted calibration voltage and the reference voltage, the deviation trend over time can be extracted, enabling the model to be updated online to adapt to the evolution of drift characteristics caused by long-term factors such as equipment aging and environmental changes. To avoid the loss of historical knowledge caused by conventional incremental learning, this invention introduces a forgetting suppression mechanism in model updates. By distinguishing between model parameters that contribute stably over a long period and those sensitive to the current environment, the model can retain existing effective experience while promptly incorporating new drift patterns. When extreme deviations occur during operation, the system can trigger the incremental learning mechanism to perform deeper parameter updates, thereby ensuring that the model can quickly recover accuracy under sudden environmental changes.

[0034] In conjunction with the above process, the system provides adaptively calibrated voltage data to the device control unit during the output stage, enabling downstream maximum power point tracking, electric vehicle charging control, and grid protection strategies to operate based on more accurate voltage information, thereby improving the overall stability and operating efficiency of the system.

[0035] For example, in scenarios where photovoltaic inverters operate for extended periods, the system can first acquire millisecond-level dynamic voltage changes via a voltage acquisition link, while simultaneously collecting temperature, humidity, and device aging information. The modeling module learns the voltage drift patterns under the combined effects of temperature and aging based on historical data. When increased solar radiation causes a rapid rise in the inverter's internal temperature, the calibration vector generated through adaptive calibration can correct the voltage deviation caused by drift in real time. When a sudden temperature and humidity anomaly occurs in the inverter's operating environment, and the calibration confidence level drops below the threshold, the system can automatically initiate online updates. If the temperature remains high and the aging process deepens, the model is further updated through forgetting suppression and incremental learning mechanisms to adapt to the new drift pattern. Ultimately, the corrected voltage data is used in the inverter's maximum power point tracking algorithm, thus maintaining high power generation efficiency even under high-temperature aging conditions.

[0036] In a preferred embodiment of the present invention, the neural network model can be a feedforward neural network, which consists of multiple layers of neurons. The input layer is used to receive the environmental-voltage feature vector, the hidden layer is used to extract nonlinear feature relationships, and the output layer is used to generate voltage drift prediction values. Through layer-by-layer weighted calculations and activation function operations, the nonlinear relationship between multiple factors such as temperature, humidity, and aging degree and voltage drift can be fitted.

[0037] In another preferred embodiment, the neural network model can employ a recurrent neural network, which is particularly suitable for processing time-series data on the evolution of voltage and environmental parameters over time. By introducing a time memory unit, long-term drift trends can be captured, and historical dependencies can be accumulated during continuous sampling. For example, in a photovoltaic power plant, the temperature rise trend of equipment over several hours can be effectively modeled using a recurrent neural network.

[0038] In another preferred embodiment, the neural network model can also be a convolutional neural network, used to extract local variation patterns of voltage data and environmental parameters. The convolutional layers can identify voltage fluctuations over a short period, thereby improving the model's adaptability to rapidly fluctuating environments.

[0039] Furthermore, to improve the robustness of the model in complex scenarios, a combined structure can be adopted. This involves using a convolutional neural network as a front-end feature extractor to capture local patterns, and then inputting the extracted features into a recurrent neural network to learn long-term temporal dependencies, thereby constructing a hybrid deep neural network model.

[0040] By selecting and combining different types of neural networks, this invention can be flexibly deployed for different application scenarios. For example, a feedforward neural network can be used for distributed inverters with small amounts of data; while for large-scale photovoltaic power plants operating continuously, a recurrent neural network or a hybrid model can be used to achieve more accurate drift prediction and calibration.

[0041] The voltage acquisition module specifically includes:

[0042] The signal input unit is used to input the analog voltage signal from the monitored power equipment into the sampling circuit. The sampling circuit includes voltage divider and isolation components to ensure that the amplitude of the input signal is within a safe range and to avoid interference between devices.

[0043] The sampling control unit is used to periodically sample the input voltage signal at a preset sampling frequency. This sampling frequency can be set according to the device characteristics; for example, a millisecond-level sampling period can be selected in photovoltaic inverter applications. This unit records a timestamp during sampling, ensuring that subsequent data maintains a consistent time reference with environmental monitoring results.

[0044] The analog-to-digital converter (ADC) unit converts analog voltage signals into digital voltage data. This conversion process is performed by a multi-bit precision ADC, ensuring that the data reflects voltage fluctuations at high resolution. To avoid the influence of quantization errors, this unit also performs oversampling and averaging filtering during the conversion process to obtain more stable raw voltage data.

[0045] The data buffer and output unit stores the raw voltage data obtained from analog-to-digital conversion in a buffer and outputs it in the sampling order, providing directly callable basic data for subsequent environmental monitoring modules and intelligent drift modeling modules.

[0046] The environmental monitoring module specifically includes:

[0047] The environmental sensing unit is used to acquire environmental information at the operating site, including temperature, humidity, air pressure, and equipment operating status parameters, such as device aging indicators or heat dissipation efficiency parameters. This unit can be implemented by combining multiple types of sensors, and time stamping is performed uniformly during data acquisition.

[0048] The environmental data processing unit is used to filter and standardize the acquired environmental signals to eliminate sensor noise and range differences. For example, for temperature signals, sudden interference can be removed using a moving average method; for humidity signals, the acquired values ​​can be normalized according to a preset range to maintain a consistent numerical scale with voltage data.

[0049] The time alignment unit is used to match the processed environmental data with the raw voltage data output by the voltage acquisition module according to the timestamp. If there are differences in sampling frequency between different signals, interpolation methods are used to fill in the time dimension, so that each voltage data point can correspond to a set of environmental parameters.

[0050] The data storage unit is used to store the aligned environmental and voltage data in key-value pairs or matrix form, thereby forming complete environmental-voltage correlation data for subsequent intelligent drift modeling modules to call.

[0051] The application output module specifically includes:

[0052] The data receiving unit receives the calibration voltage data output by the predictive calibration module and formats the data to meet the input requirements of the downstream control unit. For example, for an inverter control system, the data can be encapsulated into a real-time voltage signal frame; for an electric vehicle charging system, it is output in the data format required by the power control interface.

[0053] The application mapping unit is used to map calibration voltage data to specific control logic according to different application scenarios. For example, in photovoltaic inverters, the calibration voltage serves as the input to the maximum power point tracking algorithm to ensure that the power tracking process is based on accurate voltage values; during electric vehicle charging, the calibration voltage serves as a reference for charging power adjustment to prevent overvoltage or undervoltage operation; in grid protection scenarios, the calibration voltage serves as a real-time monitoring signal to trigger the action logic of protection devices.

[0054] The output interface unit is used to transmit calibration voltage data to the control unit via a standard communication interface. This interface can include industrial communication methods such as serial bus, Ethernet, or fieldbus to ensure the real-time performance and reliability of data transmission.

[0055] The feedback confirmation unit receives execution feedback from the control unit and compares it with the calibration voltage data to confirm whether the calibration results are being used correctly in the application scenario. If a data transmission or application anomaly is detected, this unit can generate an error signal to prompt maintenance or trigger a self-calibration mechanism.

[0056] In a preferred embodiment of the present invention, the intelligent drift modeling module includes:

[0057] The data standardization unit is used to perform time alignment and noise filtering based on the environmental-voltage correlation data to obtain standardized environmental-voltage data.

[0058] The feature vector generation unit is used to extract a multi-dimensional feature set, including time series trend features, statistical distribution features, and environmental factor correlation features, based on standardized environmental-voltage data, to obtain a feature vector.

[0059] The neural network training unit is used to train a multi-layer neural network based on feature vectors using a stepwise backpropagation and adaptive learning rate adjustment mechanism to obtain a preliminary prediction model.

[0060] The feature weight allocation unit is used to calculate the impact measure of each environmental feature on the voltage drift prediction result based on sensitivity analysis or feature importance assessment method after the initial prediction model training is completed; and to generate additional weight coefficients based on the impact measure, and to add larger weights to features with high impact to obtain the drift prediction model.

[0061] The model parameter generation unit is used to extract weight parameters and bias parameters from the drift prediction model to form prediction model parameters.

[0062] In this embodiment of the invention, by setting a data standardization unit, time alignment and noise filtering can be performed on environmental and voltage-related data, ensuring the uniformity and stability of the input data. Since the data sources collected by photovoltaic power stations or charging piles often come from multiple sensors and channels, without standardization processing, data misalignment or high-frequency noise interference can occur, leading to a decrease in model training accuracy. Preprocessing can eliminate these adverse factors before the data enters the neural network.

[0063] Furthermore, by utilizing feature vector generation units, time-series trends, statistical distributions, and environmental relevance features can be extracted from standardized data. This transforms the input data from a single dimension to a multi-dimensional feature encompassing temporal evolution, statistical regularities, and multi-factor relationships. This allows neural networks to capture, for example, voltage shifts caused by increasing temperature over time, or nonlinear drift effects resulting from the combined influence of humidity and temperature, thereby significantly improving the model's adaptability to complex environments.

[0064] Furthermore, the neural network training unit trains the multi-layer network through stepwise backpropagation and an adaptive learning rate mechanism, enabling the model to converge quickly and avoid overfitting. Compared to the traditional fixed learning rate method, the adaptive learning rate can automatically adjust the step size according to the error convergence, thereby accelerating the convergence speed when the environment changes drastically, while ensuring the accuracy of prediction when the environment is stable.

[0065] Furthermore, the feature weight allocation unit can further identify which environmental features have a greater impact on voltage drift and assign them higher weights during modeling. For example, when temperature changes have a significantly higher explanatory power for voltage drift than humidity, the system automatically increases the weight of the temperature feature, making the model more reliant on key features during calibration and reducing the interference of redundant inputs on the results.

[0066] Finally, the model parameter generation unit extracts the optimized model into a parameter set for subsequent prediction and calibration modules to use, thereby ensuring that the calibration process is based on the latest model with reasonable weight allocation and achieves more accurate real-time compensation.

[0067] The data standardization unit specifically includes:

[0068] The time synchronization processing unit is used to compare and align environmental parameters and voltage data from different sources according to the sampling timestamps. When there are inconsistencies in sampling frequencies, interpolation methods are used to fill in missing points or high-frequency data is downsampled to ensure that each set of voltage data corresponds to a complete set of environmental parameters.

[0069] The noise suppression processing unit is used to filter the aligned data to eliminate high-frequency noise and outlier interference. Filtering methods can include moving average filtering or median filtering. For example, when a temperature sensor experiences a sudden jump, filtering can smooth it out and avoid interfering with the modeling process.

[0070] The normalization unit is used to transform data of different dimensions into a uniform numerical range. For example, it transforms temperature, humidity, and voltage amplitude into the range between zero and one, so that the neural network can maintain the balance of each feature during training and avoid the model from being dominated by some feature values ​​that are too large.

[0071] Finally, the standardized output unit outputs the processed environmental-voltage data in matrix or tensor form as input for subsequent feature vector generation and neural network training.

[0072] The neural network training unit specifically includes:

[0073] The training data preparation unit is used to divide the standardized feature vectors into a training set and a validation set. The training set is used to update the model parameters, and the validation set is used to monitor the generalization performance of the model and avoid overfitting.

[0074] The forward propagation unit is used to input feature vectors into the neural network and calculate the output prediction value layer by layer. In this process, the weighted summation of neurons and the activation function operation are completed through digital processing to obtain the predicted value of voltage drift.

[0075] The error calculation unit compares the predicted value with the target value and calculates the error index. This index can be measured by converting the difference between the predicted value and the actual measured voltage into a numerical value, such as calculating the average difference or the square of the error, thereby quantifying the model's prediction accuracy.

[0076] The backpropagation unit is used to propagate gradient information layer by layer according to the error index and adjust the weights and biases in the neural network. During this process, a stepwise update method is used, meaning that a portion of the parameters are updated after each batch of data is processed, so that the model gradually converges to the optimal solution.

[0077] The learning rate adjustment unit dynamically adjusts the learning rate during training. When the error decreases rapidly, the learning step size is increased to accelerate convergence; when the error decreases slowly, the learning step size is decreased to improve accuracy. This adaptive mechanism helps avoid training stagnation or oscillation.

[0078] Finally, the trained neural network forms a preliminary prediction model and is saved for subsequent parameter generation and drift calibration.

[0079] The model parameter generation unit specifically includes:

[0080] The parameter extraction unit is used to extract the weight matrices and bias vectors of each layer after the neural network has been trained. The weight parameters represent the degree of influence of the input features on the output results, and the bias parameters are used to adjust the baseline position of the output results.

[0081] The parameter archiving unit is used to uniformly encode and store the extracted weights and biases, ensuring that they can be correctly identified in different operating environments or subsequent calls. For example, in cross-platform deployment, the parameters can be stored as standardized data files for direct loading in the prediction and calibration module.

[0082] The parameter optimization unit is used to compress and filter the extracted parameters, removing redundant or invalid connection weights to reduce model computation and improve the efficiency of the real-time calibration process. During this process, weights with values ​​close to zero can be removed by setting a threshold to simplify the model structure.

[0083] Finally, the parameter output unit packages the organized weights and bias parameters into a set of prediction model parameters, which serves as the core basis for the subsequent prediction and calibration module to perform calibration vector inference.

[0084] In a preferred embodiment of the present invention, the prediction calibration module includes:

[0085] The calibration vector inference unit is used to input the original voltage data and its corresponding environmental features into the drift prediction model according to the prediction model parameters, generate voltage offset through forward calculation, and vectorize the voltage offset to form a calibration vector.

[0086] The dynamic correction unit is used to correct the original voltage data point by point according to the calibration vector. During the correction process, the correction amplitude is adaptively adjusted according to the rate of change of the voltage signal to form preliminary calibration voltage data.

[0087] The residual calculation unit is used to calculate the residual index and form residual data based on the difference between the preliminary predicted calibration voltage data and the predicted voltage data output by the drift prediction model.

[0088] The uncertainty assessment unit is used to calculate the confidence level of the calibration result based on the residual data. When the confidence level is lower than the preset confidence threshold, a retraining trigger signal is generated and transmitted to the closed-loop self-learning module to start the retraining process of the drift prediction model. When the confidence level meets the requirements, the final predicted calibration voltage data is output.

[0089] In this embodiment of the invention, the calibration vector inference unit can input the original voltage data and environmental features into the drift prediction model according to the prediction model parameters, obtain the voltage offset through forward calculation, and further vectorize it to form a calibration vector. This process can transform the drift pattern learned in the model into actual correction instructions, ensuring that each voltage data point can receive targeted compensation.

[0090] Furthermore, the dynamic correction unit corrects the original voltage data point by point according to the calibration vector and adaptively adjusts the correction magnitude based on the rate of change of the voltage signal. This allows the system to apply a stronger correction for rapid response during rapid voltage fluctuations, while reducing the correction magnitude to avoid over-adjustment during slow voltage changes. For example, during electric vehicle charging, if the vehicle suddenly switches from low-power to high-power charging, causing increased voltage fluctuations, the dynamic correction unit can immediately increase the calibration ratio to ensure the monitoring signal remains stable.

[0091] Furthermore, the residual calculation unit can calculate the difference between the initial calibration voltage data and the model's predicted output, forming residual data to measure the calibration accuracy of the prediction model. If the residual value remains high, it indicates that the model has a bias in capturing drift, thus providing a basis for subsequent retraining.

[0092] Finally, the uncertainty assessment unit calculates the confidence level based on the residual data. When the confidence level falls below a threshold, the model retraining process is immediately triggered; when the confidence level meets the requirements, the final predicted calibration voltage data is output. This mechanism ensures the reliability of the model calibration results and avoids calibration failure due to model aging or sudden environmental changes.

[0093] The calibration vector inference unit specifically includes:

[0094] The data input unit is used to combine the raw voltage data output by the voltage acquisition module with the corresponding environmental parameters provided by the environmental monitoring module into an input vector, and then send it along with the prediction model parameters into the drift prediction model.

[0095] The forward computation unit is used to calculate the weighted sum of the activation outputs of neurons layer by layer within the drift prediction model, thereby obtaining the predicted voltage offset. Here, the weighted sum refers to multiplying the input features by their corresponding weights, summing the results, adding the bias values, and then processing the result through an activation function to obtain the output.

[0096] The offset generation unit converts the output of the forward operation into a voltage offset, which characterizes the estimated difference between the original voltage signal and the true value. The magnitude of the offset directly reflects the degree of drift in the sensor output caused by environmental influences.

[0097] The vectorization processing unit processes the voltage offset and expands it into a calibration vector. This processing may include adding time-related correction factors or weighting coefficients related to multi-dimensional environmental parameters to the offset, enabling the calibration vector to adapt to voltage compensation requirements under different time periods and environmental conditions.

[0098] Finally, the output unit provides the generated calibration vector to the dynamic correction unit to perform point-by-point correction of the original voltage data in subsequent steps.

[0099] The residual calculation unit specifically includes:

[0100] The data acquisition unit receives the preliminary predicted calibration voltage data generated by the dynamic correction unit and simultaneously obtains the predicted voltage output from the drift prediction model. Both are time-stamped to ensure the comparability of residual calculations.

[0101] The difference calculation unit compares the preliminary predicted calibration voltage data with the predicted voltage data point by point to obtain the difference value at each moment. The sign of the difference value indicates whether the calibration result is an overestimation or an underestimation, and the magnitude of the value reflects the magnitude of the deviation.

[0102] The residual evaluation unit is used to statistically process the variance values ​​to obtain residual indices. Residual indices can include the mean deviation, the variance of the variance values, and the maximum range of the variance values, comprehensively reflecting the degree of consistency between the calibration results and the model output. For example, when the average level of the variance values ​​is close to zero and the fluctuation range is small, it indicates that the model predictions are basically consistent with the actual calibration results.

[0103] The residual storage unit stores the calculated residual indices and difference sequences for later retrieval during uncertainty assessment or closed-loop self-learning. This residual data provides a quantitative basis for determining whether the model needs updating.

[0104] In a preferred embodiment of the present invention, the closed-loop self-learning module includes:

[0105] The deviation extraction unit is used to calculate the periodic deviation based on the predicted calibration voltage data and the reference voltage data of the actual operation of the equipment, and obtain the deviation sequence.

[0106] The trend analysis unit is used to extract the deviation trend based on the deviation sequence using the moving average method, and to generate deviation trend data.

[0107] The online update unit is used to iteratively adjust the drift prediction model based on the deviation trend data to obtain the updated drift prediction model, wherein the adjustment magnitude is proportional to the deviation trend.

[0108] The forgetting suppression unit is used to distinguish and mark stable weights and sensitive weights in the updated drift prediction model based on the weight change rate and long-term prediction contribution. Stable weights are retained, and sensitive weights are replaced according to a preset replacement rule to obtain a drift prediction model that retains long-term knowledge.

[0109] The evolution control unit is used to determine whether to trigger retraining based on the overall deviation performance of the drift prediction model that maintains long-term knowledge during continuous operation. When triggered, it calls the incremental learning algorithm to update the model parameters, forming the evolved drift prediction model. The updated prediction model parameters are output for the prediction calibration module and the application output module to use.

[0110] In this embodiment of the invention, the deviation extraction unit periodically compares the predicted calibration voltage with the actual reference voltage to obtain a deviation sequence. This process allows the system to rely not only on the results of the prediction model itself, but also on real-world performance for monitoring, thereby ensuring the actual effectiveness of the model.

[0111] Furthermore, the trend analysis unit uses methods such as moving averages to extract the long-term trend of the deviation, effectively filtering out the impact of short-term fluctuations and retaining only the overall trend of long-term offset. In this way, even if abnormal situations such as short-term cloud cover occur in a photovoltaic power station, it will not lead to frequent model updates, but will instead focus on long-term drift issues.

[0112] Furthermore, the online update unit can iteratively adjust the model based on the deviation trend, and the adjustment magnitude is proportional to the deviation trend, thereby achieving gradual correction. This mechanism ensures the model's adaptability to long-term drift while avoiding oscillations in prediction results caused by excessive one-time adjustments.

[0113] Furthermore, the forgetting suppression unit distinguishes between stable weights and sensitive weights during the update process, retaining stable weights and updating sensitive weights according to a preset replacement rule. This mechanism prevents the model from losing long-accumulated knowledge when adapting to new environments, thus effectively avoiding "catastrophic forgetting." For example, when the device operates in different seasons, the model can retain effective experience from the previous season while gradually introducing new environmental patterns.

[0114] Furthermore, the evolution control unit determines whether to trigger retraining based on the overall deviation performance, and employs incremental learning methods to evolve the model when triggered. This not only avoids frequent full retraining but also ensures that model parameters are continuously optimized over time, maintaining long-term adaptability. For example, in distributed power plants, the model can achieve continuous evolution by gradually introducing new operational data, thereby maintaining a stable and high level of monitoring accuracy.

[0115] The deviation extraction unit specifically includes:

[0116] The data collection subunit is used to receive the calibration voltage data output by the predictive calibration module and the reference voltage signal collected during equipment operation. The reference voltage signal can be derived from a calibrated standard voltage source or a long-term stable control measurement device.

[0117] The time alignment subunit is used to synchronize the two types of voltage data according to timestamps, ensuring that the calibration voltage and reference voltage at the same moment correspond one-to-one. If there is an inconsistency in sampling frequencies, it is adjusted through interpolation or downsampling methods.

[0118] The difference generation subunit is used to calculate the difference between the calibration voltage and the reference voltage point by point, obtaining periodic deviation data. The sign of the difference value reflects whether the calibration voltage is too high or too low relative to the reference voltage, and the magnitude of the difference value reflects the actual offset.

[0119] The data output sub-unit is used to store the continuous difference value sequence as a deviation sequence for subsequent trend analysis and model update.

[0120] The trend analysis unit specifically includes:

[0121] The sequence input subunit is used to receive the deviation sequence generated by the deviation extraction unit.

[0122] The smoothing subunit is used to set a sliding window on the deviation sequence and average the difference values ​​within the window, thereby eliminating the interference of short-term fluctuations on trend analysis. In this way, the resulting trend curve is smoother and can reflect the overall direction of long-term deviation changes.

[0123] The trend calculation subunit is used to extract the direction and rate of change of the deviation from the smoothed results. For example, when the deviation continues to increase over multiple periods, this subunit can mark it as an upward trend; when the deviation value gradually decreases, it is marked as a downward trend.

[0124] The result output sub-unit is used to output the calculated deviation trend in the form of a data sequence or trend vector for use by the online update unit.

[0125] The online update unit specifically includes:

[0126] The trend receiving subunit is used to receive the deviation trend data generated by the trend analysis unit and input it into the update logic as the basis for model updates.

[0127] The update strategy generation sub-unit is used to set the adjustment range of model parameters based on the magnitude and direction of the deviation trend. When the deviation trend continues to increase, the update range of model parameters is increased; when the deviation trend decreases, the update range of parameters is decreased to ensure the smoothness of the adjustment.

[0128] The parameter correction subunit is used to iterate through the weights and biases layer by layer within the drift prediction model and correct their values ​​according to the update strategy. The correction method adopts a stepwise cumulative form, that is, in each iteration, an adjustment amount proportional to the deviation trend is applied, so that the model gradually approaches the optimal state.

[0129] The update result output sub-unit is used to save the corrected drift prediction model as an updated model and provide it to the subsequent forgetting suppression unit and evolution control unit for use.

[0130] In a preferred embodiment of the present invention, the feature vector generation unit includes:

[0131] The time segmentation unit is used to divide the time series into sliding windows based on standardized environmental-voltage data to generate windowed sequence data;

[0132] The trend extraction unit is used to calculate the direction and magnitude of change of the sequence within the window based on the windowed sequence data, thereby forming time series trend features;

[0133] The distribution analysis unit is used to extract statistical distribution features, including mean, variance, skewness, and kurtosis, based on the time series trend characteristics to obtain distribution description data.

[0134] The correlation modeling unit is used to extract the correlation characteristics between environmental factors and voltage drift based on the distribution description data, using mutual information measurement or correlation coefficient analysis, to form environmental correlation data.

[0135] The feature combination unit is used to concatenate time series trend features, distribution description data, and environmental correlation data to obtain a multi-dimensional feature vector, which is then used by the neural network training unit.

[0136] In this embodiment of the invention, by dividing standardized environmental and voltage data into sliding windows using time segmentation units, long-term series can be transformed into multiple short-window data blocks, enabling the system to capture voltage change trends at a smaller time granularity. This processing method avoids the feature dilution problem caused by excessively long overall series, making it easier for the model to identify instantaneous change patterns.

[0137] Furthermore, the trend extraction unit analyzes the windowed data to obtain the direction and magnitude of voltage changes within the window, thus forming a time-series trend feature. This feature can reflect the upward or downward trend of voltage over a short period of time, providing a basis for the dynamic evolution of the model. For example, in a photovoltaic system, voltage changes rapidly with solar radiation intensity, and the trend feature can accurately describe the pattern of gradual voltage increase in the early morning.

[0138] Furthermore, the distribution analysis unit extracts statistical features such as mean, variance, skewness, and kurtosis from the trend characteristics, enabling the model to not only understand the direction of voltage changes but also identify their stability and degree of fluctuation. For example, a large variance indicates significant voltage fluctuations over a short period, allowing the model to assign higher weights for correction in subsequent predictions.

[0139] Furthermore, the correlation modeling unit can analyze the correlation between environmental factors and voltage drift based on distribution characteristics and generate correlation data. For example, in hot weather, the correlation between temperature and voltage drift is significantly enhanced, and this unit can automatically capture and reflect this relationship, thereby giving the model stronger explanatory power.

[0140] Finally, the feature combination unit concatenates trend features, distribution features, and correlation features to form a complete multi-dimensional feature vector, which is then input into the neural network. This process makes the model's input dimensions more comprehensive, avoiding biases caused by single features, thereby improving the accuracy of prediction and calibration.

[0141] The correlation modeling unit specifically includes:

[0142] The data preparation subunit receives the distribution description data output by the distribution analysis unit and simultaneously acquires the set of environmental parameters provided by the environmental monitoring module. Environmental parameters may include factors affecting voltage drift, such as temperature, humidity, air pressure, and degree of aging.

[0143] The correlation calculation subunit is used to analyze the numerical relationship between environmental parameters and voltage drift one by one. During the calculation, the dependence of each environmental parameter on voltage drift is first evaluated, for example, by statistically analyzing the changes in voltage drift distribution under different values. Then, the degree of linear or nonlinear correlation between the two is calculated; the higher the correlation, the stronger the explanatory power of that environmental parameter for the drift results.

[0144] The feature extraction subunit is used to transform the calculated correlation results into feature values, such as correlation strength levels or dependency scores. To avoid excessively large numerical differences affecting subsequent processing, these feature values ​​are normalized to a uniform range so that they can be combined with other types of features.

[0145] The feature output subunit is used to output the processed relevance features in vector or matrix form as environmental relevance data for use by the feature combination unit.

[0146] The feature combination unit specifically includes:

[0147] The data access subunit is used to receive time series trend features generated by the trend extraction unit, distribution description data output by the distribution analysis unit, and environmental correlation data generated by the correlation modeling unit.

[0148] The feature alignment subunit is used to uniformly process features from different sources across both the time dimension and the numerical scale. For example, when the time windows of trend features and distribution features are inconsistent, interpolation methods are used to extend them to the same time point; when the numerical ranges of features differ, normalization is used to bring them into a unified range.

[0149] The feature splicing subunit is used to concatenate and splice time trend features, distribution features, and correlation features in a predetermined order to form a complete feature vector. This vector not only contains time series information and statistical patterns, but also incorporates the influence of environmental parameters on drift, thus comprehensively reflecting the causes of voltage drift.

[0150] The output interface subunit is used to convert the formed multi-dimensional feature vector into a data format that the neural network can directly accept, and then pass it to the neural network training unit for subsequent model training and prediction capability optimization.

[0151] In a preferred embodiment of the present invention, the feature weight allocation unit includes:

[0152] The sensitivity detection unit is used to apply perturbations to the input features based on the preliminary prediction model and calculate the output difference to obtain the sensitivity index of each environmental feature.

[0153] The importance calculation unit is used to replace a single feature based on the sensitivity index and calculate the percentage decrease in prediction accuracy to form a feature importance score.

[0154] The weight mapping unit is used to generate additional weight coefficients based on feature importance scores and to adjust the original weights of the preliminary prediction model to obtain weighted weight parameters.

[0155] The model correction unit is used to optimize and update the drift prediction model based on the weighted parameters to obtain a drift prediction model with higher feature attention.

[0156] In this embodiment of the invention, the sensitivity detection unit can obtain the sensitivity index of each environmental feature to the prediction result by perturbing the input features and observing the output deviation. This can identify which features have a greater impact on the model results, providing a basis for subsequent feature selection and optimization.

[0157] Furthermore, the importance calculation unit performs permutation tests on features based on sensitivity indicators and calculates the percentage decrease in prediction accuracy to generate a feature importance score. In this way, the contribution of different environmental features can be objectively quantified. For example, if the prediction accuracy drops significantly after temperature is permuted, it indicates that temperature plays a central role in drift prediction.

[0158] Furthermore, the weight mapping unit can convert feature importance scores into additional weight coefficients, which are then combined with the model's original weights to obtain weighted parameters. This process ensures that the neural network not only relies on the weights automatically learned during training but can also be corrected based on the results of subsequent interpretive analysis, making the model's decision-making basis more reasonable.

[0159] Furthermore, the model correction unit optimizes and updates the model based on the weighted parameters, thereby making the model pay more attention to the features that have the greatest impact on voltage drift and reducing the interference of redundant features on the prediction. For example, in areas where humidity remains stable for a long period of time, the model will automatically reduce the weight of humidity features to reduce unnecessary computational costs and improve calibration efficiency.

[0160] The sensitivity detection unit specifically includes:

[0161] The input preparation subunit is used to feed the feature vector as input into the preliminary prediction model and record the model output results under normal input conditions.

[0162] The perturbation generation subunit is used to apply small changes to a single environmental feature, such as increasing or decreasing the temperature feature by a preset percentage near its original value.

[0163] The output comparison subunit is used to calculate the model's predicted output before and after the feature perturbation, and obtain the difference value between the two outputs. The larger the difference value, the more significant the impact of the environmental feature on the model's prediction.

[0164] The sensitivity assessment subunit is used to standardize the above difference values ​​and generate a sensitivity index, which represents the sensitivity of the feature in the prediction model.

[0165] The importance calculation unit specifically includes:

[0166] The feature permutation subunit is used to replace a certain environmental feature in the feature vector with random noise or irrelevant values ​​in order to disrupt the dependency between the feature and the output.

[0167] The accuracy monitoring subunit is used to rerun the prediction model and calculate the overall prediction accuracy after feature permutation. This accuracy is measured by comparing the differences between the model output and the target data, such as the average difference between the predicted and actual values.

[0168] The score generation sub-unit compares the replaced prediction accuracy with the original prediction accuracy and calculates the percentage decrease. The larger the percentage decrease, the greater the contribution of the feature to the model's prediction, thus forming the corresponding importance score.

[0169] The output sorting sub-unit is used to summarize the importance scores of all environmental features and generate a feature importance matrix, providing input for subsequent weight mapping.

[0170] The weight mapping unit specifically includes:

[0171] The weight generation subunit is used to normalize the feature importance scores into additional weight coefficients, so that the contributions of different features can be compared within a unified numerical range.

[0172] The mapping execution subunit multiplies the additional weight coefficients by the original weights in the model to obtain new weight parameters. In this way, the weights of high-contribution features are amplified, while the weights of low-contribution features are reduced, thereby enhancing the model's focus on key features.

[0173] The result cache subunit is used to store the weighted weight parameters as a parameter set for use by the model correction unit.

[0174] The model correction unit specifically includes:

[0175] The parameter loading sub-unit is used to import the weighted weight parameters output by the weight mapping unit into the drift prediction model.

[0176] The model update subunit is used to replace and correct the connection weights and biases of the model layer by layer after the new weight parameters are imported, so that it has higher feature selectivity while keeping the overall network structure unchanged.

[0177] The validation and evaluation subunit is used to test the updated model using the validation dataset and monitor whether its prediction accuracy has improved. If the validation accuracy does not meet expectations, iterative optimization can be performed by readjusting the additional weight coefficients.

[0178] Finally, the updated model is saved as a new drift prediction model and provided to the prediction calibration module for use in order to achieve higher accuracy voltage drift calibration.

[0179] In a preferred embodiment of the present invention, the dynamic correction unit includes:

[0180] The rate of change calculation unit is used to calculate the voltage rate of change based on the difference between adjacent sampling points of the original voltage data and generate rate of change data.

[0181] The fluctuation identification unit is used to divide the voltage signal into fast fluctuation range and slow change range based on the rate of change data, and obtain the range classification result;

[0182] The correction and adjustment unit is used to adaptively adjust the correction ratio of the calibration vector based on the interval classification results, increasing the correction magnitude in the rapidly fluctuating interval and decreasing the correction magnitude in the slowly changing interval, thereby generating the adjusted data.

[0183] The cumulative compensation unit is used to gradually apply compensation to the long-term offset trend based on the adjusted correction data, wherein the single compensation magnitude is limited to a preset proportion of the cumulative offset, and a smooth correction result is generated.

[0184] The output fusion unit is used to superimpose the smoothing correction results onto the original voltage data to form preliminary calibration voltage data.

[0185] In this embodiment of the invention, the rate of change calculation unit can calculate the rate of change of voltage based on the difference between adjacent sampling points. This allows for real-time monitoring of voltage fluctuations, providing a basis for subsequent classification and correction. For example, in the event of a sudden load change, the rate of change will increase instantaneously, and the system can immediately identify this situation.

[0186] Furthermore, the fluctuation identification unit divides the voltage signal into fast fluctuation ranges and slow change ranges based on the rate of change data. Through this distinction, the system can adopt different correction strategies for different voltage change scenarios. For example, when the power grid fluctuates frequently, the system enters the fast range mode to improve correction sensitivity; when the power grid is stable, it enters the slow range mode to avoid over-adjustment.

[0187] Furthermore, the correction and adjustment unit adjusts the calibration vector proportionally based on the interval classification results, enabling the correction magnitude to adaptively change with the voltage variation rate. This approach effectively improves calibration flexibility and avoids under-correction or over-correction problems caused by a single correction ratio.

[0188] Furthermore, the cumulative compensation unit applies the correction gradually and proportionally to the long-term offset trend, thereby achieving smooth error correction. For example, when a photovoltaic inverter has been operating for more than a year, sensor aging may cause the voltage to remain low. The cumulative compensation unit will gradually increase the correction value to bring the final result back to accuracy, rather than making a large adjustment all at once that could cause system oscillation.

[0189] Finally, the output fusion unit combines the smoothing correction result with the original voltage data to generate preliminary calibration voltage data, ensuring the continuity and stability of the output signal. This not only improves the accuracy of voltage monitoring but also ensures that the voltage data acquired by the downstream control unit during decision-making is stable and reliable.

[0190] The fluctuation recognition unit specifically includes:

[0191] The threshold setting subunit is used to preset a rate of change threshold based on the device's operating characteristics or historical sampling data, in order to distinguish the fluctuation patterns of the voltage signal.

[0192] The interval discrimination subunit is used to compare the rate of change data with a preset threshold. When the rate of change is greater than the threshold, the corresponding sampling point is marked as a fast fluctuation interval; when the rate of change is less than or equal to the threshold, it is marked as a slow change interval.

[0193] The interval aggregation subunit is used to merge the discrimination results of adjacent sampling points to form a continuous interval label sequence, thereby obtaining the interval classification result of the voltage signal.

[0194] The classification output sub-unit is used to output and store the interval classification results for later use in the correction and adjustment unit.

[0195] The correction and adjustment unit specifically includes:

[0196] The proportional setting subunit is used to set the correction proportional coefficient for the rapid fluctuation range and the slow change range respectively, wherein the correction proportional for the rapid fluctuation range is greater than the correction proportional for the slow change range.

[0197] The proportional application subunit is used to multiply the value of the calibration vector by the corresponding proportional coefficient in different intervals of the voltage signal, thereby forming an intervalized correction amount.

[0198] The result generation sub-unit is used to synthesize the correction amount with the original voltage data to obtain the adjusted correction data.

[0199] The cumulative compensation unit specifically includes:

[0200] The trend monitoring subunit is used to detect the average offset direction and amount of voltage data over a longer time window, thereby determining whether there is a long-term trend error.

[0201] The compensation calculation subunit is used to scale the cumulative offset according to a preset ratio to obtain the compensation magnitude for a single operation. This ratio is usually less than one, indicating that each correction only compensates for a portion of the cumulative offset, thus ensuring the gradual nature of the compensation.

[0202] The compensation application subunit is used to gradually superimpose the calculated compensation magnitude onto the adjusted correction data during continuous sampling, thereby forming a continuous and smooth compensation process.

[0203] The result storage subunit is used to save the compensated voltage data as a smoothing correction result for the output fusion unit to call.

[0204] The output fusion unit specifically includes:

[0205] The data access subunit is used to receive the smoothing correction results output by the cumulative compensation unit and the raw voltage data output by the voltage acquisition module.

[0206] The superposition subunit is used to add the smoothing correction result to the original voltage data point by point to form a calibrated voltage signal. This superposition method can compensate for the error caused by drift while ensuring the integrity of the original voltage waveform.

[0207] The result output subunit is used to output the preliminary calibration voltage data in time series form and provide it to the residual calculation unit for subsequent comparison and evaluation.

[0208] In a preferred embodiment of the present invention, the uncertainty assessment unit includes:

[0209] The residual analysis unit is used to calculate the statistical distribution characteristics of the residuals based on the residual data, including the mean, variance, and distribution range, to obtain the uncertainty index.

[0210] The confidence level generation unit is used to compare the variance of the residuals with a preset reference value based on the uncertainty index, and to count the proportion of the residuals falling into a preset tolerance range, using the proportion as the confidence level of the calibration result.

[0211] The trigger control unit generates a retraining trigger signal and transmits it to the closed-loop self-learning module when the confidence level is lower than the preset confidence threshold; when the confidence level is higher than the preset confidence threshold requirement, it outputs the final predicted calibration voltage data.

[0212] In this embodiment of the invention, the residual analysis unit can perform statistical analysis on the residuals between the preliminary calibration voltage data and the model prediction output to obtain uncertainty indicators such as mean, variance, and distribution range. This process can comprehensively reflect the fluctuation of calibration accuracy and provide a reliable basis for subsequent confidence level calculations.

[0213] Furthermore, the confidence level generation unit can use an uncertainty index to compare the variance of the residuals with a reference threshold and calculate the proportion of residuals falling within a set allowable range, thereby obtaining the confidence level of the calibration result. When the residual variance is small and most data points fall within the range, the confidence level is calculated to be high, indicating that the calibration result is highly reliable.

[0214] Furthermore, the trigger control unit determines subsequent operations based on the comparison between the confidence level and the threshold. When the confidence level is below the threshold, the system generates a retraining trigger signal and transmits it to the closed-loop self-learning module, thereby enabling the prediction model to be retrained and updated. When the confidence level is above the threshold, the system directly outputs the predicted calibration voltage data for subsequent control. This mechanism ensures the reliability of voltage calibration and prevents the model from continuing to output results when it fails. For example, in extreme weather or when long-term operation causes sensor aging, this unit can promptly detect a decrease in the confidence level and trigger retraining to maintain the accuracy of system monitoring.

[0215] The preset allowable range specifically includes:

[0216] The upper and lower limits of the residual error can be set according to the application scenario of the system. For example, in a photovoltaic inverter, the allowable range of voltage residual error can be set to between 0.5% and 1% of the rated voltage.

[0217] During system operation, the allowable range is dynamically adjusted based on environmental conditions or long-term operating results. When it is found that the residuals are generally too small or too large over a long period of time, the range will be narrowed or widened accordingly.

[0218] In a preferred embodiment of the present invention, the forgetting inhibition unit includes:

[0219] The parameter monitoring unit is used to record the rate of change of each parameter in continuous training iterations during the drift prediction model update process, and obtain the rate of change data.

[0220] The contribution assessment unit is used to calculate the contribution of each parameter to the long-term forecast performance based on the rate of change data and historical forecast accuracy, and to obtain the contribution index.

[0221] The weight protection unit is used to mark parameters with high contribution and low change rate as stable weights based on the contribution index, and to apply regularization constraints during the update process of the drift prediction model to limit its adjustment range.

[0222] The weight replacement unit is used to mark parameters with low contribution and high change rate as sensitive weights according to the contribution index, and replace them according to the preset replacement rules during the update process of the drift prediction model. The replacement rules include selecting sensitive weights that exceed the change threshold based on the change rate sorting.

[0223] The weight update execution unit is used to comprehensively call the processing results of the weight protection unit and the weight replacement unit in the overall update of the drift prediction model, so as to achieve the parallel execution of the preservation of stable weights and the replacement of sensitive weights, and obtain a drift prediction model that retains long-term knowledge.

[0224] In this embodiment of the invention, the parameter monitoring unit can record the rate of change of weight parameters in real time during model updates, thereby identifying which parameters fluctuate significantly during training iterations. This provides a quantitative basis for subsequently distinguishing between stable and unstable knowledge.

[0225] Furthermore, the contribution evaluation unit combines historical prediction accuracy to calculate the contribution of each parameter, obtaining an importance index for each weight in long-term prediction performance. In this way, the system can identify which weights play a significant role in maintaining the model's long-term stable predictions.

[0226] Furthermore, the weight protection unit marks weights with high contribution and low rate of change as stable weights based on contribution indicators, and applies regularization constraints during the update process to limit their adjustment range. This mechanism ensures that the model does not lose previously accumulated important knowledge when adapting to new environments.

[0227] Meanwhile, the weight replacement unit can identify weights with low contribution and high change rate, and perform replacement operations according to preset replacement rules. The replacement rules include sorting sensitive weights by change rate and selecting those exceeding a threshold for updating, thereby ensuring that the replacement range is controllable.

[0228] Ultimately, the weight update execution unit integrates the results of weight protection and weight replacement into the model update process, ensuring that stable weights are preserved and sensitive weights are appropriately replaced. This results in an updated model that retains long-term knowledge while adapting to new environments. For example, when a photovoltaic power station operates across seasons, the model can maintain existing knowledge of voltage drift under high summer temperatures while gradually introducing new drift patterns under low winter temperatures, achieving a balance between long-term and short-term knowledge.

[0229] The parameter monitoring unit specifically includes:

[0230] The data capture subunit is used to read the weight and bias parameters of the neural network layer by layer during model training or updates.

[0231] The change detection subunit compares the numerical differences of the same parameter between two iterations and calculates the magnitude of the change. This calculation is expressed as the ratio of the difference to the original value, without the need for a formula.

[0232] The results summary sub-unit is used to record the changes in each parameter as rate of change data and store them with the parameter index as the key value, providing a basis for subsequent contribution evaluation.

[0233] The contribution evaluation unit specifically includes:

[0234] The historical comparison subunit is used to retrieve historical prediction results and actual results data saved during system operation to evaluate the model's performance at different time periods.

[0235] The indicator calculation subunit is used to perform correlation analysis between the rate of change of each parameter and its stability with historical prediction results. When a parameter maintains a low rate of change and high accuracy in long-term prediction, its contribution is evaluated as high; when the parameter changes frequently and the prediction effect is limited, its contribution is evaluated as low.

[0236] The result output sub-unit is used to store the calculated contribution index and parameter index in correspondence, providing a basis for subsequent weight protection and replacement operations.

[0237] The weight protection unit specifically includes:

[0238] The stable filtering sub-unit is used to filter parameters from the parameter set whose contribution index is higher than a preset threshold and whose rate of change is lower than a preset threshold.

[0239] The constraint generation sub-unit is used to generate update constraints for these labeled stable weights. These constraints ensure that the parameters can only be adjusted slightly during iterative updates, without significant fluctuations.

[0240] The constraint execution subunit is used to apply constraints to the update process of stable weights during model training and updates, so that their values ​​are restricted during optimization, thereby ensuring that long-term effective knowledge is preserved.

[0241] The weight replacement unit specifically includes:

[0242] The sensitive filtering sub-unit is used to identify parameters in the contribution index that are below a preset threshold and have a change rate higher than a preset threshold, and mark them as sensitive weights.

[0243] The replacement strategy generation sub-unit is used to select sensitive weights according to replacement rules. Replacement rules may include sorting by the rate of change and selecting those exceeding a threshold, or limiting the replacement ratio to no more than a certain percentage of the total number of parameters.

[0244] The replacement execution subunit is used to replace sensitive weights with new parameter values ​​obtained from retraining when the model is updated, so that the model can adapt to the new operating environment and avoid overall performance degradation.

[0245] In a preferred embodiment of the present invention, the evolution control unit includes:

[0246] The deviation acquisition unit is used to periodically acquire the difference between the predicted calibration voltage and the reference voltage to generate deviation sequence data;

[0247] The trend detection unit is used to extract the deviation trend based on the deviation sequence data and determine whether the deviation exceeds the preset deviation threshold multiple times in a row, thereby generating a trigger judgment result.

[0248] The incremental learning unit is used to perform incremental learning on the drift prediction model when the trigger judgment result meets the evolution conditions. New data is gradually introduced while maintaining the original knowledge to obtain the evolved drift prediction model.

[0249] The parameter output unit is used to extract model parameters from the evolved drift prediction model and provide them to the prediction calibration module and the application output module.

[0250] In this embodiment of the invention, the deviation acquisition unit can periodically acquire the difference between the predicted calibration voltage and the reference voltage, and generate deviation sequence data, thereby providing the system with the ability to continuously monitor the long-term operating status.

[0251] Furthermore, the trend detection unit extracts the trend of deviation changes based on the deviation sequence and determines whether there are multiple instances of exceeding the threshold consecutively. When a continuous deviation exceeding the threshold is detected, it indicates that the existing model is insufficient to cope with environmental changes and needs to enter the evolutionary process.

[0252] Furthermore, the incremental learning unit performs incremental learning on the drift prediction model when the triggering conditions are met, gradually introducing new running data into the model without disrupting the existing knowledge structure. Compared to full retraining, incremental learning can significantly reduce computational overhead and prevent the model from forgetting old knowledge when learning new data.

[0253] Furthermore, after incremental learning is complete, the parameter output unit extracts the evolved model parameters and passes them to the prediction calibration module and the application output module, ensuring that subsequent voltage calibration and equipment control operate based on the latest evolved model. For example, in a distributed photovoltaic power station, as environmental conditions gradually change over several years of operation, the evolution control unit can ensure continuous model updates, keeping the inverter's maximum power point tracking highly efficient, thereby improving overall power generation efficiency.

[0254] The deviation acquisition unit specifically includes:

[0255] The data synchronization subunit is used to receive the predicted calibration voltage data and the reference voltage signal respectively, and to align the timestamps of the two to ensure that the data at the same sampling time can correspond one-to-one.

[0256] The difference calculation subunit compares the predicted calibration voltage with the reference voltage at each moment to obtain the instantaneous deviation value. The sign of the deviation value indicates whether the predicted voltage is overestimated or underestimated, and the magnitude reflects the difference between the two.

[0257] The sequence construction subunit is used to continuously record multiple deviation values ​​and form a deviation sequence within a set sampling period, which serves as the input data for the trend detection unit.

[0258] The trend detection unit specifically includes:

[0259] The smoothing subunit is used to perform a sliding window average on the biased sequence to remove short-term random fluctuations and obtain a smoother sequence.

[0260] The trend analysis sub-unit is used to identify the direction and rate of change of deviation in the smoothed sequence. When the deviation value gradually increases over multiple periods, it is determined to be an upward trend; when the deviation gradually decreases, it is determined to be a downward trend.

[0261] The threshold judgment subunit is used to compare the deviation trend with the preset threshold. When the deviation exceeds the threshold for multiple consecutive periods, a judgment result is generated indicating that the trigger condition is met.

[0262] The incremental learning unit specifically includes:

[0263] The data preparation subunit is used to extract the latest voltage and environmental data from the system operation records when the triggering conditions are met, and integrate them with historical training data.

[0264] The incremental update sub-unit is used to input new data in small batches. Each batch of data will apply a small update to the model parameters to ensure that the model retains old knowledge while gradually adapting to the new environment.

[0265] The knowledge preservation subunit is used to impose constraints on important weights during parameter updates to prevent them from being adjusted drastically, thereby avoiding the forgetting of previously learned effective rules.

[0266] The resulting sub-unit is used to label the updated drift prediction model as the evolved model, which has better adaptability to the new environment.

[0267] The parameter output unit specifically includes:

[0268] The parameter extraction subunit is used to read the weight parameters and bias parameters of the neural network layer by layer after the model evolution is completed, to ensure that the data is complete and error-free.

[0269] The parameter packaging subunit is used to uniformly encode the extracted parameters and save them as a callable data file or memory data block for cross-module transmission.

[0270] The interface output subunit is used to send parameters to the prediction calibration module and the application output module through a standardized interface, so that subsequent calibration and control processes can run based on the latest model parameters.

[0271] The status confirmation subunit is used to receive feedback from downstream modules to confirm that the parameters have been loaded and applied correctly; if an abnormality is found, it returns an error message to prompt reloading or updating again.

[0272] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An intelligent voltage monitoring system with adaptive calibration function, characterized in that, The system includes: The voltage acquisition module is used to receive real-time voltage signals from the monitored power equipment, perform analog-to-digital conversion, and generate raw voltage data. The environmental monitoring module is used to synchronously collect operating environmental parameters and store them in time correspondence with the raw voltage data to form environmental-voltage correlation data. The intelligent drift modeling module is used to train environmental-voltage correlation data based on a neural network model, learn the nonlinear drift law under multi-factor coupling conditions, construct a drift prediction model, and generate prediction model parameters representing voltage drift trends. The prediction calibration module is used to generate a calibration vector based on the prediction model parameters and adaptively adjust the calibration amplitude according to the changing characteristics of the voltage signal; and calculate the calibration confidence level based on the residual between the calibration result and the model predicted voltage value. When the confidence level is lower than the preset confidence threshold, a retraining trigger signal is sent to the closed-loop self-learning module to generate prediction calibration voltage data. The closed-loop self-learning module is used to extract the deviation trend based on the periodic deviation between the predicted calibration voltage and the reference voltage, and update the drift prediction model online accordingly. The closed-loop self-learning module also includes a forgetting suppression mechanism to maintain long-term effective knowledge, and triggers incremental learning to achieve model evolution and update when the deviation trend continues to exceed the limit. The application output module is used to provide the predicted calibration voltage data to the control unit of the power equipment to support intelligent decision-making for photovoltaic inverter maximum power point tracking control, electric vehicle charging power regulation, or grid safety protection. The intelligent drift modeling module includes: The data standardization unit is used to perform time alignment and noise filtering based on the environmental-voltage correlation data to obtain standardized environmental-voltage data. The feature vector generation unit is used to extract a multi-dimensional feature set, including time series trend features, statistical distribution features, and environmental factor correlation features, based on standardized environmental-voltage data, to obtain a feature vector. The neural network training unit is used to train a multi-layer neural network based on feature vectors using a stepwise backpropagation and adaptive learning rate adjustment mechanism to obtain a preliminary prediction model. The feature weight allocation unit is used to calculate the impact measure of each environmental feature on the voltage drift prediction result based on sensitivity analysis or feature importance assessment method after the initial prediction model training is completed; and to generate additional weight coefficients based on the impact measure, and to add larger weights to features with high impact to obtain the drift prediction model. The model parameter generation unit is used to extract weight parameters and bias parameters from the drift prediction model to form prediction model parameters.

2. The intelligent voltage monitoring system with adaptive calibration function according to claim 1, characterized in that, The prediction calibration module includes: The calibration vector inference unit is used to input the original voltage data and its corresponding environmental features into the drift prediction model according to the prediction model parameters, generate voltage offset through forward calculation, and vectorize the voltage offset to form a calibration vector. The dynamic correction unit is used to correct the original voltage data point by point according to the calibration vector. During the correction process, the correction amplitude is adaptively adjusted according to the rate of change of the voltage signal to form preliminary calibration voltage data. The residual calculation unit is used to calculate the residual index and form residual data based on the difference between the preliminary predicted calibration voltage data and the predicted voltage data output by the drift prediction model. The uncertainty assessment unit is used to calculate the confidence level of the calibration result based on the residual data. When the confidence level is lower than the preset confidence threshold, a retraining trigger signal is generated and transmitted to the closed-loop self-learning module to start the retraining process of the drift prediction model. When the confidence level meets the requirements, the final predicted calibration voltage data is output.

3. The intelligent voltage monitoring system with adaptive calibration function according to claim 1, characterized in that, The closed-loop self-learning module includes: The deviation extraction unit is used to calculate the periodic deviation based on the predicted calibration voltage data and the reference voltage data of the actual operation of the equipment, and obtain the deviation sequence. The trend analysis unit is used to extract the deviation trend based on the deviation sequence using the moving average method, and to generate deviation trend data. The online update unit is used to iteratively adjust the drift prediction model based on the deviation trend data to obtain the updated drift prediction model, wherein the adjustment magnitude is proportional to the deviation trend. The forgetting suppression unit is used to distinguish and mark stable weights and sensitive weights in the updated drift prediction model based on the weight change rate and long-term prediction contribution. Stable weights are retained, and sensitive weights are replaced according to a preset replacement rule to obtain a drift prediction model that retains long-term knowledge. The evolution control unit is used to determine whether to trigger retraining based on the overall deviation performance of the drift prediction model that maintains long-term knowledge during continuous operation. When triggered, it calls the incremental learning algorithm to update the model parameters, forming the evolved drift prediction model. The updated prediction model parameters are output for the prediction calibration module and the application output module to use.

4. The intelligent voltage monitoring system with adaptive calibration function according to claim 1, characterized in that, The feature vector generation unit includes: The time segmentation unit is used to divide the time series into sliding windows based on standardized environmental-voltage data to generate windowed sequence data; The trend extraction unit is used to calculate the direction and magnitude of change of the sequence within the window based on the windowed sequence data, thereby forming time series trend features; The distribution analysis unit is used to extract statistical distribution features, including mean, variance, skewness, and kurtosis, based on the time series trend characteristics to obtain distribution description data. The correlation modeling unit is used to extract the correlation characteristics between environmental factors and voltage drift based on the distribution description data, using mutual information measurement or correlation coefficient analysis, to form environmental correlation data. The feature combination unit is used to concatenate time series trend features, distribution description data, and environmental correlation data to obtain a multi-dimensional feature vector, which is then used by the neural network training unit.

5. The intelligent voltage monitoring system with adaptive calibration function according to claim 1, characterized in that, The feature weight allocation unit includes: The sensitivity detection unit is used to apply perturbations to the input features based on the preliminary prediction model and calculate the output difference to obtain the sensitivity index of each environmental feature. The importance calculation unit is used to replace a single feature based on the sensitivity index and calculate the percentage decrease in prediction accuracy to form a feature importance score. The weight mapping unit is used to generate additional weight coefficients based on feature importance scores and to adjust the original weights of the preliminary prediction model to obtain weighted weight parameters. The model correction unit is used to optimize and update the drift prediction model based on the weighted parameters to obtain a drift prediction model with higher feature attention.

6. The intelligent voltage monitoring system with adaptive calibration function according to claim 2, characterized in that, The dynamic correction unit includes: The rate of change calculation unit is used to calculate the voltage rate of change based on the difference between adjacent sampling points of the original voltage data and generate rate of change data. The fluctuation identification unit is used to divide the voltage signal into fast fluctuation range and slow change range based on the rate of change data, and obtain the range classification result; The correction and adjustment unit is used to adaptively adjust the correction ratio of the calibration vector based on the interval classification results, increasing the correction magnitude in the rapidly fluctuating interval and decreasing the correction magnitude in the slowly changing interval, thereby generating the adjusted data. The cumulative compensation unit is used to gradually apply compensation to the long-term offset trend based on the adjusted correction data, wherein the single compensation magnitude is limited to a preset proportion of the cumulative offset, and a smooth correction result is generated. The output fusion unit is used to superimpose the smoothing correction results onto the original voltage data to form preliminary calibration voltage data.

7. The intelligent voltage monitoring system with adaptive calibration function according to claim 2, characterized in that, The uncertainty assessment unit includes: The residual analysis unit is used to calculate the statistical distribution characteristics of the residuals based on the residual data, including the mean, variance, and distribution range, to obtain the uncertainty index. The confidence level generation unit is used to compare the variance of the residuals with a preset reference value based on the uncertainty index, and to count the proportion of the residuals falling into a preset tolerance range, using the proportion as the confidence level of the calibration result. The trigger control unit generates a retraining trigger signal and transmits it to the closed-loop self-learning module when the confidence level is lower than the preset confidence threshold; when the confidence level is higher than the preset confidence threshold requirement, it outputs the final predicted calibration voltage data.

8. The intelligent voltage monitoring system with adaptive calibration function according to claim 3, characterized in that, The forgetting inhibition unit includes: The parameter monitoring unit is used to record the rate of change of each parameter in continuous training iterations during the drift prediction model update process, and obtain the rate of change data. The contribution assessment unit is used to calculate the contribution of each parameter to the long-term forecast performance based on the rate of change data and historical forecast accuracy, and to obtain the contribution index. The weight protection unit is used to mark parameters with high contribution and low change rate as stable weights based on the contribution index, and to apply regularization constraints during the update process of the drift prediction model to limit its adjustment range. The weight replacement unit is used to mark parameters with low contribution and high change rate as sensitive weights according to the contribution index, and replace them according to the preset replacement rules during the update process of the drift prediction model. The replacement rules include selecting sensitive weights that exceed the change threshold based on the change rate sorting. The weight update execution unit is used to comprehensively call the processing results of the weight protection unit and the weight replacement unit in the overall update of the drift prediction model, so as to achieve the parallel execution of the preservation of stable weights and the replacement of sensitive weights, and obtain a drift prediction model that retains long-term knowledge.

9. The intelligent voltage monitoring system with adaptive calibration function according to claim 3, characterized in that, The evolution control unit includes: The deviation acquisition unit is used to periodically acquire the difference between the predicted calibration voltage and the reference voltage to generate deviation sequence data; The trend detection unit is used to extract the deviation trend based on the deviation sequence data and determine whether the deviation exceeds the preset deviation threshold multiple times in a row, thereby generating a trigger judgment result. The incremental learning unit is used to perform incremental learning on the drift prediction model when the trigger judgment result meets the evolution conditions. New data is gradually introduced while maintaining the original knowledge to obtain the evolved drift prediction model. The parameter output unit is used to extract model parameters from the evolved drift prediction model and provide them to the prediction calibration module and the application output module.

Images