High-voltage SVG adaptive voltage control method based on deep learning

Abstract

The invention discloses a high-voltage SVG adaptive voltage control method based on deep learning, and belongs to the technical field of power equipment, and the method comprises the following steps: S1, multi-source data collection and preprocessing: the S1, multi-source data collection and preprocessing comprises a data collection module and a data preprocessing module, the data acquisition module comprises a high-precision sensor network and acquires various data in real time; the data preprocessing module comprises missing value filling through a sliding window method, wavelet transform denoising, multi-dimensional feature proof construction and normalization processing; s2, constructing a deep learning model, wherein the construction of the deep learning model in S2 comprises a model framework module, a network parameter module and a training strategy module. According to the method, the nonlinear characteristics of the power grid are captured through the deep learning model, the method can adapt to dynamic scenes such as power grid topological structure sudden change and new energy output intermittence in real time, and in the power grid with the new energy permeability exceeding 30%, the voltage regulation error can be reduced to + / -0.5% and is remarkably superior to + / -2% of a traditional method.

Description

Technical Field

[0001] This invention belongs to the field of power equipment technology, specifically a high-voltage SVG adaptive voltage control method based on deep learning. Background Technology

[0002] In the power system, with the large-scale integration of new energy sources and the increasing complexity of the power system, the proportion of new energy in the power system is constantly rising. At the same time, the complexity of the power system is also growing exponentially, leading to increasingly significant voltage fluctuations in the power grid. On the one hand, new energy power generation, especially wind and solar power, is significantly constrained by natural factors, exhibiting highly intermittent and uncertain output. Wind power fluctuations depend on the rapidly changing wind speed and direction, while solar power is affected by meteorological conditions such as sunlight intensity and cloud cover. This randomness in new energy output makes it difficult to accurately control the power balance of the power grid, thus causing large voltage fluctuations. On the other hand, the characteristics of the load side have also changed significantly. Taking electric vehicles as an example, their widespread adoption and disorderly charging behavior can create impact loads on the power grid. A large number of concentrated charging sessions in a short period of time can cause a sudden drop in local grid voltage. Similarly, the start-up and shutdown of large equipment in industrial production and the widespread application of nonlinear loads also exacerbate the randomness and complexity of the load, further disrupting the stability of the power grid voltage.

[0003] In traditional power systems, voltage fluctuations mainly originate from normal load variations and output adjustments by a few generator units. The grid topology is relatively stable, and various parameters change gradually and are highly predictable. Static var generators (SVG), as the core device for dynamic reactive power compensation, directly affect grid voltage stability through their control accuracy and response speed. Traditional SVG control methods are mostly based on fixed-parameter PID control or static reactive power compensation strategies, making it difficult to adapt to the strong random disturbances and multivariate coupling characteristics under scenarios with a high proportion of renewable energy integration. Especially in grid environments where wind and solar power output is increasingly intermittent and load fluctuations are frequent, the operating environment of the power system becomes more complex than ever before.

[0004] This complexity manifests itself primarily in the following aspects: With the continuous development of the power grid, frequent operations such as line switching and transformer tap adjustments lead to constant changes in the grid topology. This dynamic change makes traditional control methods based on fixed topologies difficult to adapt. Wind power, photovoltaic power, and other new energy power generation are greatly affected by natural conditions, resulting in significant intermittency and uncertainty in output. Random fluctuations in loads such as electric vehicle charging and industrial production make the load characteristics of the power grid difficult to predict. The integration of power electronic devices such as new energy generation units, flexible AC transmission systems (FACTS), and energy storage devices, while improving the flexibility of the power grid, also introduces new problems such as harmonics and phase jumps. Therefore, a high-voltage SVG adaptive voltage control method based on deep learning is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based adaptive voltage control method for high-voltage SVG to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution, comprising the following steps: S1: Multi-source data acquisition and preprocessing, which includes a data acquisition module and a data preprocessing module. The data acquisition module includes a high-precision sensor network and real-time acquisition of various types of data. The data preprocessing module includes sliding window method to fill missing values, wavelet transform denoising, construction of multi-dimensional feature evidence and normalization processing. S2: Deep learning model construction, which includes a model architecture module, a network parameter module, and a training strategy module. The model architecture module adopts a Long Short-Term Memory Network (LSTM-Attention) model with a hybrid attention mechanism. The input layer is multi-dimensional temporal features, the hidden layer includes spatial attention module and temporal attention module, and the output layer is SVG control instructions. The network parameter module includes setting parameters for the number of LSTM layers and the number of hidden units. The training strategy module includes offline training and online updating. S3: Adaptive control strategy generation, which includes a model output control command module, a safety constraint verification module, and a reinforcement learning feedback loop module. The model output control command module generates the PWM signal of the SVG, the safety constraint verification module verifies the SVG capacity limit constraint, and the reinforcement learning feedback loop module dynamically adjusts the model weights. As a further preferred embodiment of this technical solution: the data acquisition module acquires power grid status data, SVG operation data, environmental data, and device status data; As a further preferred embodiment of this technical solution: the power grid status data includes bus voltage amplitude, phase angle, and frequency; the SVG operating data includes output reactive power, DC side capacitor voltage, and IGBT temperature; the environmental data includes new energy output forecast, load fluctuation curve, and meteorological parameters; and the equipment status data includes SVG cooling fan speed and cooling system pressure. As a further preferred embodiment of this technical solution: the data preprocessing module uses the sliding window method to fill in missing values, uses wavelet transform to denoise, constructs a multi-dimensional feature matrix containing time-series features, statistical features and frequency domain features, and performs normalization processing on the feature matrix to eliminate dimensional differences; As a further preferred embodiment of this technical solution: in the sliding window method, the window length = 1s, the step size = 0.1s, the time series features and statistical features include mean, variance, and peak value, and the frequency domain features are the main frequency components after FFT transformation; As a further preferred embodiment of this technical solution: the model architecture module captures the spatial correlation between different sensor data and the long-term dependence of time series data; the SVG control commands include reactive power compensation amount and modulation ratio. As a further preferred embodiment of this technical solution: the network parameter module has 3 LSTM layers and 128 hidden units; 4 spatial attention heads and 8 temporal attention heads; the loss function is weighted MSE, where the reactive power compensation error weight is 0.7 and the harmonic suppression error weight is 0.3. As a further preferred embodiment of this technical solution: the offline training in the training and measurement module is based on historical operating data to construct a training set, and adopts the Adam optimizer and dynamic learning rate adjustment mechanism. The online update in the training and measurement module uses an incremental learning mechanism to fine-tune the model parameters with new data every 5 minutes to adapt to changes in the power grid topology. As a further preferred embodiment of this technical solution: after the model output control command module is verified by safety constraints, it generates a pulse width modulation signal for the SVG. The safety constraint module includes: SVG capacity limiting, IGBT temperature rise protection, and bus voltage safety boundary. As a further preferred embodiment of this technical solution: the reinforcement learning feedback loop module dynamically adjusts the model weights according to the voltage regulation effect to achieve closed-loop optimization of the control strategy. The reward function is designed as: R=-α|ΔV|-β|ΔQ|, where: ΔV is the voltage deviation, ΔQ is the reactive power compensation error, α=0.6, β=0.4.

[0007] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention captures the nonlinear characteristics of the power grid through a deep learning model, and can adapt to dynamic scenarios such as sudden changes in the power grid topology and intermittent power output of new energy sources in real time. In power grids with a new energy penetration rate of more than 30%, the voltage regulation error can be reduced to ±0.5%, which is significantly better than ±2% of the traditional method.

[0008] 2. This invention utilizes lightweight model deployment and TensorRT to accelerate inference, achieving a model inference time of ≤5ms and a total control command generation time of <15ms. This significantly reduces the time compared to traditional methods, enabling rapid response during periods of severe load fluctuations and avoiding the risk of voltage exceeding limits.

[0009] 3. Through loss function design and reinforcement learning feedback loop, this invention can simultaneously optimize reactive power compensation efficiency (≥98%), harmonic suppression rate (THD≤2%), and system stability. In the scenario of fluctuating new energy output, the harmonic suppression rate is significantly improved compared with the traditional method (THD>5%).

[0010] 4. The reinforcement learning feedback loop of this invention dynamically adjusts the model weights based on the voltage regulation effect, realizing online adaptive adjustment of the control strategy. In scenarios such as grid fault ride-through and intermittent power output of new energy sources, it can continuously optimize control performance and improve the robustness of the power grid. Attached Figure Description

[0011] Figure 1 This is an overall flowchart of a high-voltage SVG adaptive voltage control method based on deep learning according to the present invention. Figure 2 This is a flowchart of multi-source data acquisition and preprocessing for a high-voltage SVG adaptive voltage control method based on deep learning, as described in this invention. Figure 3 This is a flowchart of the deep learning model construction process for a high-voltage SVG adaptive voltage control method based on deep learning, as described in this invention. Figure 4 This is a flowchart illustrating the adaptive control strategy generation process of a high-voltage SVG adaptive voltage control method based on deep learning, as described in this invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] A Static Var Generator (SVG) is a dynamic reactive power compensation device based on power electronics technology (typically using fully controlled devices such as IGBTs). It can quickly, continuously, and accurately generate or absorb reactive power, thereby maintaining system voltage stability, improving power factor, and enhancing power quality.

[0014] Accordingly, SVG can be applied in multiple fields such as power systems, new energy, metallurgy, manufacturing, and mining. The following section primarily introduces the applications of SVG in power systems and new energy.

[0015] In power systems, SVG can be applied to various scenarios such as voltage stabilization control, wind farm / photovoltaic power plant grid connection points, substations, and power quality management nodes to improve the stability and reliability of power systems.

[0016] Specifically, when applied to voltage stabilization control, SVG can be installed at the end of long-distance transmission lines, load centers, or weak nodes to provide dynamic reactive power support, suppress voltage drops or rises, and improve system stability and transmission capacity.

[0017] When SVG is applied at the grid connection point of wind farms / photovoltaic power plants, the intermittent and fluctuating nature of renewable energy output leads to voltage fluctuations at the grid connection point. SVG can provide rapid reactive power regulation, meeting the voltage and power factor requirements of grid connection guidelines, and is key to solving reactive power demand during "low-voltage / high-voltage" periods.

[0018] When SVG is used in substations, it can compensate for reactive power losses within the station and maintain the bus voltage within the acceptable range.

[0019] When SVG is applied at power quality management nodes, it can be installed upstream of sensitive loads or at points with weak power quality to comprehensively solve problems such as voltage fluctuations, flicker, and imbalance.

[0020] In the field of new energy, SVG can be applied to various scenarios such as wind farms, photovoltaic power stations and energy storage power stations to stabilize voltage and ensure the stability and reliability of the power system connected to the power plant.

[0021] Specifically, when SVG is applied in wind farms, the wind turbines (especially doubly fed and full-power converter models) absorb reactive power, and wind speed changes cause power output fluctuations, leading to voltage fluctuations and power factor issues at the grid connection point. Therefore, SVG is a standard configuration that meets grid specifications (such as providing reactive power support during low voltage ride-through) and ensures voltage stability within the power plant.

[0022] When SVG is applied in photovoltaic power plants, photovoltaic inverters typically operate at unity power factor, but variations in solar irradiance and cloud cover cause rapid fluctuations in power output, affecting the grid connection voltage. Correspondingly, SVG provides dynamic reactive power compensation to stabilize voltage and meet grid connection requirements.

[0023] When SVG is used in energy storage power stations, reactive power support is needed to maintain voltage stability when the energy storage converter switches between charging and discharging modes.

[0024] Therefore, SVG (Static Var Generator) is primarily used in environments with rapid and significant fluctuations in reactive power demand, and where high requirements are placed on voltage stability, power factor, or power quality. From heavy industries with severe impact loads (such as electric arc furnaces and rolling mills) to the grid connection of fluctuating renewable energy power plants (wind power and solar power), and to power system nodes and critical facilities sensitive to power quality (data centers and rail transit), SVG is a key technological tool for solving dynamic reactive power problems in modern power systems and improving grid stability and power quality. Its application scenarios are becoming increasingly widespread with the increasing penetration of electronic loads and renewable energy sources.

[0025] Please see Figures 1-4 As shown, this invention provides a high-voltage SVG adaptive voltage control method based on deep learning, comprising the following steps: S1: Multi-source data acquisition and preprocessing. S1: Multi-source data acquisition and preprocessing includes a data acquisition module and a data preprocessing module. The data acquisition module includes a high-precision sensor network and real-time acquisition of various types of data. The data preprocessing module includes sliding window method to fill missing values, wavelet transform denoising, construction of multi-dimensional feature evidence and normalization processing. The high-precision sensor network architecture also includes power grid status monitoring, SVG body monitoring and environmental perception system. Among them, the power grid status monitoring deploys μPMU device to achieve 10kHz synchronous sampling, adopts IEEE C37.118.2 protocol for data synchronization, and the voltage / current measurement accuracy reaches 0.2%. This high-precision monitoring method provides a reliable basis for real-time understanding of the power grid operation status. For example, when a power grid fault or load change occurs, it can quickly capture subtle changes in voltage and current, providing timely and accurate information for subsequent control decisions. Fiber optic current transformers (FOCTs) are configured to detect the DC component of 500kV lines. SVG body monitoring incorporates NTC thermistor arrays embedded in the IGBT modules, constructing a three-dimensional temperature field model. Real-time monitoring of IGBT temperature effectively prevents device damage caused by overheating, ensuring stable SVG operation. Simultaneously, a hybrid energy storage system combining film and electrolytic capacitors is configured for the DC capacitors, with pressure sensors monitoring capacitor deformation. This helps to promptly detect potential capacitor faults and take early maintenance measures. Furthermore, the environmental sensing system integrates lidar and a micro-weather station to acquire real-time wind speed / direction data within 10km of the wind farm. Kalman filtering enables 5-second ultra-short-term power prediction. For power grids integrating wind power and other new energy sources, this allows for early prediction of power change trends, enabling the SVG to adjust its control strategy in advance and better cope with the intermittency and fluctuations in new energy output. S2: Deep Learning Model Construction. S2: Deep learning model construction includes a model architecture module, a network parameter module, and a training strategy module. The model architecture module uses a Long Short-Term Memory (LSTM-Attention) network model with a hybrid attention mechanism. The input layer contains multi-dimensional temporal features, the hidden layers include spatial and temporal attention modules, and the output layer is SVG control commands. The network parameter module includes setting the number of LSTM layers and hidden units. The training strategy module includes offline training and online updates. The input layer receives 128-dimensional preprocessed features and uses BatchNormalization to accelerate training convergence. The spatial attention module designs a 4-head self-attention mechanism, compressing the Query / Key / Value matrix to 64 dimensions. Spatial feature recalibration is achieved through deformable convolutions, with a 3×3 kernel size and dynamically adjusted dilation rate. The inter-attention module employs an 8-head causal convolutional attention mechanism, with a mask matrix to ensure no leakage of future information. A time-encoding mechanism is introduced, adding sine / cosine positional encoding to the input sequence. The hidden layer consists of 3 LSTM units, each with 128 cells, using a Zoneout strategy to randomly skip cell state updates. The retention rate is set to 0.15 to effectively alleviate overfitting. The fully connected layer in the output layer generates the reactive power compensation command Qref and modulation ratio M. The activation function uses the Tanh function to limit the output range, enabling the model to fully capture the long-term dependencies of time-series data. For example, considering the impact of changes in grid load over a period of time on the current voltage state, the reactive power output of the SVG is reasonably adjusted. The output layer is the SVG control command, including the reactive power compensation command Qref and modulation ratio M. The activation function uses the Tanh function to limit the output range, ensuring that the control command is within a reasonable range and guaranteeing the safe and stable operation of the SVG. S3: Adaptive Control Strategy Generation. S3: Adaptive Control Strategy Generation includes a model output control command module, a safety constraint verification module, and a reinforcement learning feedback loop module. The model output control command module generates the PWM signal for the SVG. The safety constraint verification module verifies the SVG capacity limiting constraint. The reinforcement learning feedback loop module dynamically adjusts the model weights. The PWM modulation module uses a space vector modulation (SVPWM) algorithm with a switching frequency of 2.5kHz and develops carrier phase shift technology to reduce the three-phase output voltage THD to 0.8%. The safety constraint verification module includes capacity limiting, temperature rise protection, and voltage boundary. The capacity limiting Qref is limited to [0, Qmax]. Qmax is dynamically adjusted according to the SVG's heat dissipation capacity: Qmax = Qrated(1 - 0.05(T_igbt - 60)). The temperature rise protection logic activates overload protection when the predicted IGBT junction temperature exceeds 80℃, reducing the output current to 80% of the rated value within 200ms. The voltage boundary bus voltage automatically engages the dynamic braking resistor when it exceeds 1.05pu; and activates SVG reactive power compensation when it falls below 0.95pu. In this embodiment, specifically: the data acquisition module collects power grid status data, SVG operation data, environmental data, and device status data, and synchronously measures the line active power, reactive power, and RMS current value in the power grid status data through the PMU device; In this embodiment, specifically: the power grid status data includes bus voltage amplitude, phase angle, and frequency, with a sampling rate ≥10kHz and an accuracy of 0.2%. These data directly reflect the voltage quality and operating conditions of the power grid. The SVG operating data includes output reactive power, DC side capacitor voltage, and IGBT temperature. The environmental data includes predicted values ​​of new energy output, load fluctuation curves, and meteorological parameters. The equipment status data includes SVG cooling fan speed and cooling system pressure. The load fluctuation curve includes historical data plus short-term predictions with a time resolution of 1 second. This is crucial for accurately grasping the load change pattern and helps the SVG make reactive power compensation adjustments in advance. It is used to understand the operating status of the SVG in real time so that the constraints of the SVG itself can be considered during the control process. Equipment status data such as SVG cooling fan speed and cooling system pressure can be used to evaluate the operating status of the SVG auxiliary system and indirectly ensure the reliable operation of the SVG. In this embodiment, specifically: the data preprocessing module uses a sliding window method to fill in missing values, performs wavelet transform denoising, constructs a multi-dimensional feature matrix containing time-series features, statistical features, and frequency domain features, normalizes the feature matrix to eliminate dimensional differences, and uses statistical methods to detect and correct sensor noise, such as voltage spikes. Missing values ​​are then filled using linear interpolation or prediction based on historical data. The data preprocessing includes dynamic compensation for missing values, wavelet threshold denoising, feature matrix construction, and data standardization. The dynamic compensation for missing values ​​uses an improved sliding window method with a window length of 1 second (corresponding to 20 power frequency cycles for a 50Hz system) and a step size of 0.1 seconds. For data with more than 3 consecutive missing sampling points, an LSTM prediction model is used for interpolation. The model training data includes 72 hours of historical operating conditions. This approach considers both the timeliness of the data and utilizes historical data to mine potential patterns, effectively solving... To address the issue of missing data and ensure the integrity of the input data for the deep learning model, wavelet thresholding denoising employs a 5-level decomposition using the db8 wavelet basis. Soft thresholding is used for high-frequency coefficients in levels 3-5. An adaptive thresholding algorithm is developed to address IGBT switching noise, with a threshold λ = σ√(2logN) / 0.6745, where σ is the noise standard deviation and N is the signal length. This effectively removes noise interference, preserves the signal's effective features, improves data quality, and provides a clean data foundation for subsequent model training. The feature matrix construction extracts time-domain features (mean, variance, peak factor), frequency-domain features (FFT main frequency amplitude / phase), and time-frequency features (wavelet entropy), constructing a 128-dimensional feature vector. Time-domain features account for 40%, frequency-domain features 35%, and time-frequency features 25%. This feature matrix design fully leverages the characteristic information of the data in different domains, providing a more comprehensive representation of the power grid and SVG. The running status provides rich input features for deep learning models. Among them, the data standardization adopts the RobustScaler method. For outlier-sensitive data such as IGBT temperature, the median and interquartile range are used for scaling, which effectively suppresses the influence of outliers on the model, enables data of different dimensions to be processed on the same scale, accelerates model training convergence, and improves the model's generalization ability. In this embodiment, specifically: the window length = 1s, the step size = 0.1s in the sliding window method, the time series features and statistical features include mean, variance, and peak value, and the frequency domain features are the main frequency components after FFT transformation; In this embodiment, specifically: the model architecture module captures the spatial correlation between different sensor data and the long-term dependency of time series data; the SVG control command includes reactive power compensation amount and modulation ratio. In this embodiment, specifically: the number of LSTM layers in the network parameter module is 3, the number of hidden units is 128; the number of spatial attention heads is 4, the number of temporal attention heads is 8; the loss function is weighted MSE, where the reactive power compensation error weight is 0.7 and the harmonic suppression error weight is 0.3. In this embodiment, specifically: offline training in the training and measurement module constructs a training set based on historical operating data, with a sample size > 10^6, employing the Adam optimizer and a dynamic learning rate adjustment mechanism. Online updates in the training and measurement module utilize an incremental learning mechanism (learning rate = 0.001, β1 = 0.9, β2 = 0.999; when the verification loss does not decrease for 5 consecutive rounds, the learning rate decays to 0.5 times). Every 5 minutes, new data is used to fine-tune the model parameters to adapt to changes in the power grid topology, such as line switching. The Adam optimizer, with its adaptive learning rate adjustment and good convergence performance, can effectively accelerate model convergence in large-scale data training. The dynamic learning rate adjustment mechanism adjusts the learning rate in a timely manner according to changes in loss during training, improving training efficiency and model performance. This strategy enables the model to adapt to changes in the power grid topology in a timely manner, such as line switching, and continuously maintain the model's accurate tracking and control capability of the power grid operating status, ensuring that the SVG control strategy always meets the actual operating requirements of the power grid. In this embodiment, specifically: after the model output control command module undergoes safety constraint verification, it generates a pulse width modulation signal for the SVG. The safety constraint module includes: SVG capacity limiting, IGBT temperature rise protection, and bus voltage safety boundary. The SVG capacity limiting is 0≤Q≤Q_max, where Q_max is the rated capacity of the SVG. The IGBT temperature rise protection is T<85℃. The bus voltage safety boundary is 0.95pu≤V≤1.05pu. Through this high-precision PWM modulation technology, effective control of the SVG output voltage is achieved, ensuring that the output voltage quality meets the grid requirements and guaranteeing the accurate output of the SVG reactive power compensation effect. In this embodiment, specifically: the reinforcement learning feedback loop dynamically adjusts the model weights based on the voltage regulation effect, such as regulation time and overshoot, to achieve closed-loop optimization of the control strategy. The reward function is designed as: R = -α|ΔV| - β|ΔQ|, where ΔV is the voltage deviation, ΔQ is the reactive power compensation error, α = 0.6, and β = 0.4. The reinforcement learning optimization mechanism includes state space definition and action space design. The state space definition includes 12-dimensional state variables such as voltage deviation ΔV, reactive power error ΔQ, and IGBT temperature T. The action space design includes adjusting the LS... The scaling factor of the LSTM hidden layer weight matrix is ​​limited to ±5% of the action range. Through reinforcement learning, the control effect is fed back to the model to form a closed-loop optimization. The state space is defined to include 12-dimensional state variables such as voltage deviation ΔV, reactive power error ΔQ, and IGBT temperature T, which comprehensively reflects the operating status of the power grid and SVG. The action space is designed to adjust the scaling factor of the LSTM hidden layer weight matrix, with the action range limited to ±5%, so as to achieve adaptive optimization of the model and continuously improve the performance of the SVG control strategy, so that it can achieve good voltage control effect under different operating conditions.

[0026] Working principle or structural principle: In the data preprocessing stage, the sliding window method is used to fill in missing values ​​to ensure data integrity; wavelet transform is used to remove noise interference and improve data quality; a multi-dimensional feature matrix is ​​constructed to deeply integrate time-series features (such as mean, variance, peak value), statistical features (such as the cumulative effect of historical control commands), and frequency domain features (such as FFT main frequency amplitude / phase) to fully extract valuable information from the data; finally, the feature matrix is ​​normalized to eliminate dimensional differences, creating a good data environment for model training; and high-precision sensor networks are used to collect real-time power grid status data (such as bus voltage). The data includes amplitude, phase angle, and frequency; SVG operating data (such as output reactive power, DC-side capacitor voltage, and IGBT temperature); environmental data (such as predicted renewable energy output, load fluctuation curves, and meteorological parameters); and equipment status data (such as SVG cooling fan speed and cooling system pressure). This data provides comprehensive input information for the deep learning model. A sliding window method is used to fill in missing values, and wavelet transform is used to remove noise. This constructs a multi-dimensional feature set that includes time-series features (such as mean, variance, and peak value), statistical features (such as the cumulative effect of historical control commands), and frequency domain features (such as the main frequency component after FFT transformation). The matrix is ​​normalized to eliminate dimensional differences and improve model training efficiency. The model training process is divided into two stages: offline training and online update. Offline training uses massive historical operating data to build a training set and employs the Adam optimizer and dynamic learning rate adjustment mechanism to efficiently train the model, enabling it to initially grasp the operating rules of the power grid and the SVG control strategy. Online update uses an incremental learning mechanism to fine-tune the model parameters in real time using new data, allowing the model to quickly adapt to dynamic situations such as changes in power grid topology, fluctuations in new energy output, and changes in load characteristics, ensuring that the model always keeps pace with the actual power grid situation. The system maintains a high degree of consistency with actual operation. It employs a hybrid attention mechanism, Long Short-Term Memory (LSTM-Attention) network. The input layer consists of multi-dimensional temporal features, and the hidden layers include a spatial attention module (capturing spatial correlations between data from different sensors) and a temporal attention module (capturing long-term dependencies in temporal data). The output layer contains SVG control commands (such as reactive power compensation and modulation ratio). The system is configured with 3 LSTM layers, 128 hidden units, 4 spatial attention heads, and 8 temporal attention heads. The loss function is weighted MSE (with a reactive power compensation error weight of 0.7 and a harmonic suppression error weight of 0).3) Offline training builds a training set based on historical operational data, employing the Adam optimizer and a dynamic learning rate adjustment mechanism. Online updates utilize an incremental learning mechanism, fine-tuning model parameters every 5 minutes using new data to adapt to changes in the power grid topology. The deep learning model, leveraging a hybrid attention mechanism and a Long Short-Term Memory (LSTM-Attention) network architecture, accurately captures the complex spatial relationships between different sensor data and the long-term dependencies in time-series data. The input layer receives pre-processed multi-dimensional time-series features, while the spatial and temporal attention modules in the hidden layer work together to deeply extract key features from the data, laying the foundation for generating accurate SVG control commands. After safety constraint verification, the generated SVG... The pulse width modulation (PWM) signal, with safety constraints including SVG capacity limiting, IGBT temperature rise protection, and bus voltage safety boundaries, dynamically adjusts model weights based on voltage regulation effects (such as regulation time and overshoot) to achieve closed-loop optimization of the control strategy. In the closed-loop optimization stage, the reinforcement learning feedback loop dynamically adjusts model weights based on voltage regulation effects. After review by the safety constraint verification module, the control commands output by the model are converted into SVG pulse width modulation (PWM) signals. PWM modulation employs a space vector modulation (SVPWM) algorithm, combined with carrier phase shifting technology, to effectively reduce the total harmonic distortion (THD) of the three-phase output voltage, ensuring the quality of the SVG output voltage and accurately executing reactive power compensation tasks. Through a carefully designed reward function, the accuracy and response speed of voltage regulation are quantified. Combined with the definitions of state space and action space, continuous optimization of the model is achieved, constantly improving the control performance of the SVG. This ensures that the SVG maintains stable grid voltage in complex and ever-changing power system environments, guaranteeing the safe and reliable operation of the power system and effectively addressing the voltage fluctuation challenges brought about by large-scale integration of new energy sources and increased power system complexity. The reward function is designed as: R = -α|ΔV| - β|ΔQ|, where ΔV is the voltage deviation, ΔQ is the reactive power compensation error, α = 0.6, and β = 0.4. This fully considers the heat dissipation of the SVG under different operating temperatures, reasonably limits its reactive power output capacity, prevents equipment damage due to overload operation, and includes temperature rise protection. In terms of voltage boundary conditions, when the predicted junction temperature of the IGBT exceeds 80℃, the overload protection logic is activated promptly, reducing the output current to 80% of the rated value within 200ms, effectively protecting the IGBT module and extending the equipment's lifespan. During voltage boundary verification, if the bus voltage exceeds 1.05 pu, the dynamic braking resistor is automatically engaged; if it is below 0.95 pu, SVG reactive power compensation is activated to ensure the bus voltage remains within a reasonable range, guaranteeing grid voltage stability. From multi-source data acquisition and preprocessing, deep learning model construction to adaptive control strategy generation, each stage is closely linked and collaborates with each other. It fully utilizes the powerful data mining and modeling capabilities of deep learning, combined with the closed-loop optimization mechanism of reinforcement learning, to achieve precise and adaptive control of the high-voltage SVG.

[0027] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0028] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A high-voltage SVG adaptive voltage control method based on deep learning, characterized in that: Includes the following steps: S1: Multi-source data acquisition and preprocessing, which includes a data acquisition module and a data preprocessing module. The data acquisition module includes a high-precision sensor network and real-time acquisition of various types of data. The data preprocessing module includes sliding window method to fill missing values, wavelet transform denoising, construction of multi-dimensional feature evidence and normalization processing. S2: Deep learning model construction, which includes a model architecture module, a network parameter module, and a training strategy module. The model architecture module adopts a Long Short-Term Memory Network (LSTM-Attention) model with a hybrid attention mechanism. The input layer is multi-dimensional temporal features, the hidden layer includes spatial attention module and temporal attention module, and the output layer is SVG control instructions. The network parameter module includes setting parameters for the number of LSTM layers and the number of hidden units. The training strategy module includes offline training and online updating. S3: Adaptive control strategy generation, which includes a model output control command module, a safety constraint verification module, and a reinforcement learning feedback loop module. The model output control command module generates the PWM signal of the SVG, the safety constraint verification module verifies the SVG capacity limit constraint, and the reinforcement learning feedback loop module dynamically adjusts the model weights.

2. The high-voltage SVG adaptive voltage control method based on deep learning according to claim 1, characterized in that: The data acquisition module collects power grid status data, SVG operation data, environmental data, and device status data.

3. The high-voltage SVG adaptive voltage control method based on deep learning according to claim 2, characterized in that: The power grid status data includes bus voltage amplitude, phase angle, and frequency; the SVG operating data includes output reactive power, DC side capacitor voltage, and IGBT temperature; the environmental data includes new energy output forecast, load fluctuation curve, and meteorological parameters; and the equipment status data includes SVG cooling fan speed and cooling system pressure.

4. The high-voltage SVG adaptive voltage control method based on deep learning according to claim 3, characterized in that: The data preprocessing module uses a sliding window method to fill in missing values, performs wavelet transform for noise reduction, and constructs a multi-dimensional feature matrix containing temporal features, statistical features, and frequency domain features. The feature matrix is ​​normalized to eliminate dimensional differences.

5. The high-voltage SVG adaptive voltage control method based on deep learning according to claim 4, characterized in that: In the sliding window method, the window length is 1s and the step size is 0.1s. The time series features and statistical features include the mean, variance, and peak value. The frequency domain features are the main frequency components after FFT transformation.

6. The high-voltage SVG adaptive voltage control method based on deep learning according to claim 5, characterized in that: The model architecture module captures the spatial correlation between different sensor data and the long-term dependence of time-series data. The SVG control commands include reactive power compensation and modulation ratio.

7. The high-voltage SVG adaptive voltage control method based on deep learning according to claim 6, characterized in that: The network parameter module has 3 LSTM layers and 128 hidden units; 4 spatial attention heads and 8 temporal attention heads; the loss function is weighted MSE, with reactive power compensation error weighting of 0.7 and harmonic suppression error weighting of 0.

3.

8. The high-voltage SVG adaptive voltage control method based on deep learning according to claim 7, characterized in that: The offline training in the training and measurement module is based on historical operating data to build a training set, and adopts the Adam optimizer and dynamic learning rate adjustment mechanism. The online update in the training and measurement module uses an incremental learning mechanism to fine-tune the model parameters with new data every 5 minutes to adapt to changes in the power grid topology.

9. The high-voltage SVG adaptive voltage control method based on deep learning according to claim 8, characterized in that: After the model output control command module is verified by safety constraints, it generates the pulse width modulation signal of SVG. The safety constraint module includes: SVG capacity limiting, IGBT temperature rise protection, and bus voltage safety boundary.

10. The high-voltage SVG adaptive voltage control method based on deep learning according to claim 9, characterized in that: The reinforcement learning feedback loop module dynamically adjusts the model weights based on the voltage regulation effect to achieve closed-loop optimization of the control strategy. The reward function is designed as: R=-α|ΔV|-β|ΔQ|, where: ΔV is the voltage deviation, ΔQ is the reactive power compensation error, α=0.6, β=0.4.

Images