Distributed optical storage on-site autonomous reactive voltage control system and method

Abstract

This invention discloses a distributed photovoltaic-storage on-site autonomous reactive power and voltage control system and method. It achieves efficient extraction of transformer area features through a lightweight edge sensing terminal, generates dynamic reactive power adjustment commands using a two-layer reinforcement learning framework, and realizes coordinated control of photovoltaic and energy storage devices by combining a distributed consensus protocol and a priority response strategy. This invention is applicable to weak communication scenarios in medium- and low-voltage distribution networks, providing reliable voltage control technology for high-proportion renewable energy integration.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, and in particular to a distributed photovoltaic-storage local autonomous reactive voltage control system and method based on lightweight distribution area voltage situation awareness and self-learning technology. Background Technology

[0002] With the large-scale integration of distributed photovoltaic (PV) and energy storage systems (ESS) into power distribution networks, power fluctuations on both the source and load sides have intensified significantly, leading to frequent problems such as voltage exceeding limits and three-phase imbalance in distribution areas. Traditional solutions rely on high-density smart meters (AMIs) for data collection, but this method consumes a great deal of communication bandwidth and edge computing resources and is difficult to meet real-time control requirements.

[0003] Current voltage regulation mainly adopts a centralized optimization strategy at the master station, which requires coordination of multiple node devices through a remote communication system. However, this model has two significant drawbacks: first, the data transmission time between the master station and the terminal devices is prolonged, resulting in a lag in control response; second, the complexity of the centralized optimization algorithm increases exponentially with the number of nodes, making it impossible to run in real time on the limited computing resources at the edge of the distribution area.

[0004] Current reactive power compensation devices (such as SVG and capacitor banks) mostly adopt periodic fixed switching strategies, which cannot adapt to minute-level source-load fluctuations. Although distributed photovoltaic and energy storage devices have reactive power regulation potential, the lack of an effective multi-node coordination mechanism results in their dynamic reactive power support capabilities not being fully utilized. These problems severely restrict the voltage quality control efficiency in high-penetration distributed energy scenarios. Summary of the Invention

[0005] To address the above problems, this invention provides a distributed photovoltaic energy storage local autonomous reactive power and voltage control system and method.

[0006] The technical solution of this invention is: a distributed photovoltaic-storage on-site autonomous reactive power and voltage control system, comprising a lightweight edge sensing terminal, a self-learning dynamic decision-making module, and a distributed collaborative execution unit connected in sequence, wherein,

[0007] The lightweight edge sensing terminal is deployed at key nodes in the distribution area (including distribution transformers, photovoltaic inverters, energy storage converters, and load terminals) to collect voltage, current, and power data in real time at a sampling frequency of 1kHz. The data is then reduced in dimensionality using sliding window filtering (window length 100ms) and principal component analysis (PCA) algorithms to extract features such as voltage deviation and harmonic distortion rate (the compressed data dimension is ≤10 dimensions).

[0008] The self-learning dynamic decision-making module is based on a two-layer reinforcement learning framework. The upper layer is a deep Q-network (DQN) at the distribution area level, which takes into account the overall voltage deviation of the distribution area, photovoltaic power output prediction data and load trend data, and outputs a global voltage regulation target. The lower layer is a device-level distributed Actor-Critic algorithm, which generates reactive power output instructions (Q_ref) for each photovoltaic and energy storage device based on local voltage characteristics and the upper layer target.

[0009] Distributed collaborative execution units, embedded in photovoltaic inverters and energy storage converters, exchange state information (voltage, reactive power output) with neighboring nodes via a sparse communication protocol (LoRa / HPLC) and optimize the objective function based on a consensus protocol.

[0010]

[0011] In the formula, N is the total number of voltage nodes, V i Let V be the voltage at the i-th voltage node. ref I is the reference voltage for the transformer area. i For the circuit flowing through the i-th branch, R i Let be the impedance of the i-th branch;

[0012] α is the weight of the voltage deviation term, and β is the weight of the line loss term. Where α = 0.7, β = 0.3, and a priority response strategy is implemented (energy storage devices take precedence over photovoltaic inverters).

[0013] The lightweight edge sensing terminal supports multiple communication protocols such as LoRa, HPLC, and RS485, with a communication interval of 5 seconds and a communication distance of 500 meters.

[0014] The sliding window filtering algorithm filters out 50Hz power frequency interference while retaining the dynamic characteristics of voltage fluctuations (time constant ≤ 100ms).

[0015] The principal component analysis (PCA) algorithm has a cumulative variance contribution rate of ≥94%, achieving a data compression of over 80% of the original data.

[0016] The self-learning dynamic decision-making module supports incremental online learning. It embeds the pre-trained model into the edge device through transfer learning technology and updates the policy network weights every 15 minutes based on real-time data. The edge device computing resource utilization rate is ≤15%.

[0017] The decision cycle of the device-level distributed Actor-Critic algorithm is 1 second, the reactive power output adjustment accuracy is ±5kVar, the energy storage device response time is ≤1 second, and the photovoltaic inverter response time is ≤2 seconds.

[0018] The priority response strategy of the distributed collaborative execution unit includes: when the node voltage exceeds the limit...

[0019] (|V i -V ref When the value is greater than 5% per unit, the adjustment weight of the node is increased by 50%; the reactive power adjustment priority of the energy storage device is higher than that of the photovoltaic inverter, and it provides capacitive or inductive reactive power support first (adjustment range ±100kVar).

[0020] In the objective function of the consensus protocol, the weights α and β of the voltage deviation term and the line loss term can be dynamically adjusted, with initial values ​​of 0.7 and 0.3, respectively.

[0021] A method for local autonomous reactive power and voltage control of distributed photovoltaic energy storage includes the following steps:

[0022] Step 1: Lightweight Situational Awareness

[0023] The edge sensing terminal collects key node data at a frequency of 1kHz, removes noise through a 100ms sliding window filter, and uses the PCA algorithm to compress the data dimension to less than 10 dimensions to extract features such as voltage deviation and harmonic distortion rate.

[0024] Step 2: Self-learning decision generation

[0025] The upper-layer DQN network generates a global voltage regulation target based on the overall characteristics of the transformer area. The lower-layer Actor-Critic algorithm combines the local voltage status with the global target to calculate the reactive power output command (Q_ref) of each device.

[0026] Step 3: Distributed Cooperative Execution

[0027] Each photovoltaic and energy storage device exchanges status information every 5 seconds through sparse communication. Based on the consensus protocol, the objective function is optimized, and priority is given to responding to nodes that exceed voltage limits. The energy storage device and the photovoltaic inverter work together to perform reactive power regulation, thereby achieving dynamic voltage balance in the distribution area (response time ≤ 5 seconds).

[0028] In step 2, incremental online learning is supported. The pre-trained model is embedded into the edge device through transfer learning technology. The pre-trained model is trained for more than 50,000 steps based on historical data of the transformer area (including at least 3 weather scenarios such as sunny / rainy / cloudy). The policy network weights are updated every 15 minutes during online learning.

[0029] In step 3, the communication protocol is compatible with LoRa and HPLC, the data volume of a single node in a single communication is ≤100 bytes, and the communication bandwidth requirement is reduced by more than 70% compared with the traditional solution.

[0030] The present invention has the following advantages:

[0031] 1. Improved response speed: Breaking through the minute-level delay of traditional centralized control, if the photovoltaic output in the distribution area suddenly increases (change rate > 10% / second), the voltage regulation response time can be shortened to ≤ 3 seconds, effectively dealing with voltage swell / droop issues.

[0032] 2. Communication efficiency optimization: Through data compression and sparse communication, if the LoRa communication protocol is adopted, the communication bandwidth requirement can be reduced to <50kbps, which is more than 70% lower than the traditional solution. The amount of data transmitted per node per transmission is reduced by 80%, solving the network congestion caused by high-density data transmission.

[0033] 3. Enhanced operating performance:

[0034] Voltage qualification rate: In typical transformer area testing, if the frequency of operating condition changes is ≤5 times / hour, the voltage qualification rate can be increased from 88% to 99%, meeting the distribution network voltage quality standard (GB / T12325).

[0035] Network loss reduction: 12%-18% lower than centralized model predictive control (MPC). If the power factor of the distribution area load is below 0.9, the network loss reduction effect can be further improved to 20%.

[0036] Device utilization: Edge device computing resource utilization is ≤15%, and it can be optionally adapted to low-cost embedded hardware such as ARM Cortex-A53 and Raspberry Pi.

[0037] 4. Compatibility and scalability: Supports multiple types of communication protocols and photovoltaic storage equipment. If the penetration rate of new energy in the distribution area is ≤30%, this system can be deployed directly. Optionally, through parameter adaptive adjustment, it can be compatible with scenarios with higher penetration rates (such as above 50%). Attached Figure Description

[0038] Figure 1 This is the system architecture diagram of the present invention (edge ​​perception layer - self-learning decision layer - device execution layer). Detailed Implementation

[0039] like Figure 1 As shown, the distributed photovoltaic-storage on-site autonomous reactive power and voltage control system provided by this invention includes the following core modules, forming a hierarchical control architecture, specifically:

[0040] 1. Lightweight edge sensing terminal

[0041] Deployment and Data Acquisition: Edge terminals are deployed at transformers, photovoltaic inverters, energy storage converters, and load terminals in the distribution area. These terminals integrate voltage / current sensors and optional multi-protocol communication modules (supporting LoRa, HPLC, RS485, or other wireless communication protocols compliant with the IEEE 802.15.4 standard) to collect voltage, current, active power, and reactive power data in real time at a sampling frequency of 1kHz.

[0042] Data processing and feature extraction: Optionally, a 100ms sliding window filtering algorithm or wavelet transform algorithm can be used to filter out 50Hz power frequency interference while retaining the dynamic characteristics of voltage fluctuations (time constant ≤ 100ms); through principal component analysis (PCA) or other dimensionality reduction algorithms (such as LDA), the original 100+ dimensional data can be optionally compressed to within 10 dimensions (cumulative variance contribution rate ≥ 94%), and key features such as voltage deviation (per unit value, accuracy ± 0.5%) and harmonic distortion rate can be extracted, achieving a data transmission compression of more than 80%.

[0043] Communication mechanism: Supports sparse communication with a 5-second interval, with a single node transmitting ≤100 bytes of data at a time, and a communication distance covering 500 meters. If the communication environment is a high-density distribution area, the communication interval can be optionally adjusted to 2 seconds to ensure data real-time performance.

[0044] 2. Self-learning dynamic decision-making module

[0045] Two-layer reinforcement learning framework: Upper layer is a substation-level deep Q-network (DQN): Inputs include the overall voltage deviation of the substation exceeding a preset threshold (e.g., ±5% per unit value), the photovoltaic output prediction for the next 15 minutes (1kW resolution), and the load trend for the past 24 hours (1-minute sampling interval), and outputs the global voltage regulation target (reference voltage V_ref for each node).

[0046] The lower-level device-level distributed Actor-Critic algorithm generates reactive power output commands (Q_ref) based on local voltage status (real-time acquired values) and upper-level targets in a 1-second decision cycle. The adjustment range of the photovoltaic inverter is ±50kVar (response time ≤2 seconds), the adjustment range of the energy storage converter is ±100kVar (response time ≤1 second), and the adjustment accuracy is ±5kVar.

[0047] Incremental online learning: The pre-trained model is trained for more than 50,000 steps based on historical data that optionally includes at least three weather scenarios such as sunny, rainy, and cloudy. The edge device updates the policy network weights every 15 minutes through transfer learning technology. If the edge device is a low-cost ARM Cortex-A53 processor, the computing resource utilization rate is ≤15%.

[0048] 3. Distributed collaborative execution unit

[0049] Consistency Protocol Optimization: Each optical storage device exchanges voltage and reactive power output status through sparse communication. The following function is optimized with the goal of minimizing voltage deviation and line loss:

[0050]

[0051] In the formula, N is the total number of voltage nodes, V i Let V be the voltage at the i-th voltage node. ref I is the reference voltage for the transformer area. i For the circuit flowing through the i-th branch, R i Let be the impedance of the i-th branch;

[0052] The voltage deviation term has a weight of α = 0.7, and the line loss term has a weight of β = 0.3. Dynamic adjustment is optional (e.g., automatically switching the α / β values ​​according to the peak and valley periods of the load).

[0053] Priority response strategy: If a node voltage exceeding the limit is detected (|V i -V ref If the voltage exceeds 5% PU, the adjustment weight of that node is increased by 50%; the energy storage device, with its faster response speed (≤1 second), can optionally take priority over the photovoltaic inverter in performing reactive power compensation, if the voltage over-limit type is a sag (V i If the reactive power is >1.05 PU, then inductive reactive power support will be provided first; if it is a sag (V i If the reactive power is less than 0.95 pu, capacitive reactive power support will be provided first.

[0054] This invention achieves efficient extraction of transformer area features through a lightweight edge sensing terminal, generates dynamic reactive power adjustment commands using a two-layer reinforcement learning framework, and realizes collaborative control of photovoltaic and energy storage equipment by combining a distributed consensus protocol and a priority response strategy.

[0055] A distributed photovoltaic energy storage local autonomous reactive power and voltage control method includes:

[0056] Lightweight situational awareness: Edge terminals collect electrical data of key nodes at a frequency of 1kHz. After noise is removed by 100ms sliding window filtering or median filtering, the data dimension is compressed to less than 10 dimensions by PCA algorithm to extract features such as voltage deviation and harmonic distortion rate. The processing cycle is ≤2 seconds.

[0057] Self-learning decision generation: The upper-layer DQN network generates a global voltage regulation target (e.g., V) based on the overall features of the transformer area. ref If the target voltage deviates from the real-time voltage by more than 3% per unit, the lower-level Actor-Critical algorithm combines the local voltage status with the global target to calculate the reactive power output command Q for each device. refThe pre-trained model updates the policy network incrementally every 15 minutes based on real-time data, and can optionally shorten the update cycle to 5 minutes when the output of new energy sources fluctuates drastically.

[0058] Distributed collaborative execution: The photovoltaic and energy storage devices exchange status information every 5 seconds via the LoRa / HPLC protocol. If the communication link is interrupted for a short time, a local caching strategy can be optionally enabled to maintain the adjustment command of the previous cycle. The objective function is optimized based on the consensus protocol, and priority is given to responding to nodes that exceed the limit. The reactive power output is adjusted collaboratively by the energy storage and photovoltaic devices to achieve dynamic voltage balance in the distribution area. The overall response time is ≤5 seconds.

[0059] Specific examples are as follows:

[0060] I. Example 1: Deployment of a typical 10kV distribution substation

[0061] Application scenario: Urban areas with a new energy penetration rate of 25%, including 10 photovoltaic inverters (50kW capacity per unit), 2 energy storage converters (100kVar capacity per unit), and the edge terminal adopts ARM Cortex-A53 processor (1.2GHz clock speed, 512MB memory).

[0062] (I) System Hardware Configuration

[0063] Edge sensing terminal:

[0064] Sensors: voltage transformer (accuracy class 0.2), current transformer (accuracy class 0.5), sampling frequency 1kHz.

[0065] Communication modules: Optional LoRa wireless module (470MHz band, 500m communication distance) and HPLC power line carrier module (supports IEEE1901.1 protocol), with a default communication interval of 5 seconds.

[0066] Hardware model: Advantech UNO-1122G, dimensions 100×100×50mm, operating temperature -20℃~+60℃.

[0067] Photovoltaic storage equipment:

[0068] Photovoltaic inverter: Huawei SUN2000-50KTL, reactive power regulation range ±50kVar, response time ≤2 seconds. Energy storage converter: Sungrow PCS-100K, reactive power regulation range ±100kVar, response time ≤1 second, battery capacity 200kWh.

[0069] (II) Data Processing and Decision Parameters

[0070] Lightweight perception:

[0071] Sliding window filtering: 100ms window length (100 sampling points), filters out 50Hz fundamental frequency interference, and retains 2nd-5th harmonic characteristics.

[0072] PCA dimensionality reduction: Input 100-dimensional raw data (instantaneous voltage and current values), output 10-dimensional feature quantities (voltage deviation, phase imbalance, harmonic distortion rate, etc.), with a cumulative variance contribution rate of 95.2%.

[0073] Self-learning decision-making:

[0074] DQN network structure: 10-dimensional input layer → fully connected layer (64 neurons, ReLU activation) → 3-dimensional output layer (each node V_ref), target update cycle 100 steps.

[0075] The Actor-Critic algorithm: the policy network outputs Q_ref (continuous value), the critic network evaluates the state value, the learning rate is 1e-4, and the experience replay buffer capacity is 10000.

[0076] Online learning: The latest hour of data (including 300 working condition samples) is extracted every 15 minutes for transfer learning, with edge device CPU utilization at 12% to 15%.

[0077] (III) Collaborative Control Process

[0078] Voltage over-limit response (if phase A voltage > 1.05 pu is detected):

[0079] At time T0: The edge terminal detects a phase A voltage of 1.06 pu, triggering the over-limit flag.

[0080] T2 seconds: Over-limit information is broadcast via LoRa, and the energy storage converter responds first, outputting -50kVar inductive reactive power (to suppress voltage rise).

[0081] T5 seconds: The photovoltaic inverter outputs -30kVar reactive power based on the local ΔV=+0.06pu, and after coordinated adjustment, the voltage of phase A drops to 1.02pu.

[0082] Consistency protocol execution:

[0083] The objective function weights are α = 0.7 and β = 0.3. Each node swaps V every 5 seconds. i Q i The optimal Q is obtained by using the gradient descent method on the data. ref The calculation error is ≤1%.

[0084] II. Example 2: Optimization of High-Permeability Low-Pressure Substation (380V)

[0085] Application scenario: Rural transformer substation with a photovoltaic penetration rate of 35%. The load is mainly single-phase motors (power factor 0.8), and there is a three-phase imbalance problem. The communication method is RS485 (wired, distance 200 meters).

[0086] (I) Differentiated Configuration

[0087] Edge terminal:

[0088] Add a three-phase current sensor to calculate the three-phase imbalance in real time (formula: (I max -I min ) / I avg The threshold is 15% above the limit.

[0089] An optional harmonic compensation mode can be enabled. When THD > 5%, energy storage devices are prioritized to provide harmonic reactive power compensation.

[0090] Control strategy adjustment:

[0091] Priority response: If the voltage of a certain phase exceeds the limit and the three-phase imbalance is >15%, the adjustment weight of the equipment in that phase is increased by 100% (α=1.4).

[0092] The Actor-Critic algorithm takes three-phase current amplitude and phase difference as inputs and outputs phase-specific reactive power commands (Qa_ref, Qb_ref, Qc_ref) to achieve precise phase-specific adjustment.

[0093] (II) Typical Operating Condition Handling

[0094] Three-phase imbalance adjustment (B-phase voltage 0.92 pu, A / C phase voltage 1.0 pu):

[0095] At time T0: The voltage deviation of phase B is detected to be -8%, triggering phase-by-phase adjustment.

[0096] T1 second: The energy storage converter outputs +80kVar capacitive reactive power to phase B, increasing the voltage to 0.98pu.

[0097] T3 seconds: The adjacent node photovoltaic inverter adds +20kVar, and the final B-phase voltage stabilizes at 0.99pu, and the three-phase imbalance decreases from 20% to 5%.

[0098] Emergency response to communication interruption:

[0099] If the RS485 link failure lasts for more than 10 seconds, the edge terminal can optionally switch to local independent control mode: based on the last received global target (V_ref = 1.0pu), it adjusts according to a preset strategy (e.g., for every 1% voltage deviation, the photovoltaic output increases by 5kVar) until communication is restored.

[0100] III. Example 3: Robustness Test under Extreme Conditions

[0101] Test conditions: Typhoon weather caused a sudden drop in photovoltaic output (from full capacity to 20% within 10 minutes), while the load suddenly increased (+50%), simulating a voltage dip scenario (V=0.9pu).

[0102] (I) Model Adaptation Mechanism

[0103] Incremental learning trigger:

[0104] When a voltage deviation of >5% is detected for three consecutive cycles (3 seconds), the edge terminal automatically shortens the learning cycle to 5 minutes and prioritizes the collection of current operating condition data (increasing the sample weight by 50%).

[0105] Policy network update:

[0106] New emergency regulation rule: If the voltage drops and the energy storage SOC is greater than 30%, the energy storage converter will directly output +100kVar capacitive reactive power (exceeding the conventional gradient calculation).

[0107] The edge device's computing resource utilization temporarily increased to 20% for ≤15 minutes, without triggering overload protection.

[0108] (II) Test Results

[0109] Voltage recovery time: It takes 4.2 seconds from the occurrence of a voltage dip to stabilization at 0.98 pu, which is better than traditional droop control (120 seconds).

[0110] Network loss variation: The peak network loss during the adjustment process was 15kW, which decreased to 8kW after stabilization, a reduction of 40% compared to centralized control.

[0111] IV. Example 4: Communication Protocol Compatibility Verification

[0112] Test configuration: hybrid deployment of LoRa (5 nodes) and HPLC (3 nodes) to verify cross-protocol collaboration capabilities.

[0113] Data interaction process:

[0114] Edge terminals encapsulate data uniformly through a protocol conversion module (supporting Modbus / TCP). LoRa nodes transmit data at 5-second intervals, HPLC nodes at 3-second intervals, and the system automatically synchronizes timestamps.

[0115] Cooperative accuracy:

[0116] When communicating across protocols, the voltage deviation calculation error is ≤0.3% per unit, and the reactive power command consistency reaches 98%, meeting the requirements of GB / T 35727-2017 "Technical Specifications for Distributed Power Generation Access to Distribution Network".

[0117] The working principle of this invention is as follows:

[0118] 1. Lightweight edge perception technology: A combination of sliding window filtering and PCA dimensionality reduction is proposed to achieve efficient feature extraction at the edge. If the original data contains harmonic interference, the feature extraction accuracy can be improved by adding a wavelet transform preprocessing step, which significantly reduces the amount of data transmission and computational load.

[0119] 2. Two-layer reinforcement learning decision-making framework: Construct a two-layer control framework that combines regional global optimization and device-level local autonomy. If the upper-layer DQN network detects a trend shift in the voltage of the distribution area (duration > 10 minutes), it can optionally trigger the emergency adjustment mode of the lower-layer Actor-Critic algorithm to improve the autonomous decision-making ability under complex operating conditions.

[0120] 3. Distributed collaborative control mechanism: Design a sparse communication strategy and priority response rules based on a consensus protocol (such as "if the voltage exceeds the limit, the energy storage takes priority action"), and optionally support adaptive reconfiguration when a node fails (such as skipping offline nodes for collaborative computing), thus breaking through the collaborative control bottleneck in weak communication scenarios.

[0121] The embodiments described above are only for illustrating the technical solutions of the present invention, and are not intended to limit it. Although specific embodiments have been described in detail herein, those skilled in the art can still modify them or replace some technical features in an equivalent manner, and these changes do not depart from the core ideas and protection scope embodied in the embodiments of the present invention.

Claims

1. A distributed photovoltaic-storage on-site autonomous reactive power and voltage control system, characterized in that, It includes a lightweight edge-aware terminal, a self-learning dynamic decision-making module, and a distributed collaborative execution unit connected in sequence, wherein, Lightweight edge sensing terminals are deployed at key nodes in power distribution areas to collect voltage, current, and power data in real time, and perform data dimensionality reduction and feature extraction. The self-learning dynamic decision-making module is based on a two-layer reinforcement learning framework. The upper layer is a deep Q-network at the distribution area level, which takes into account the overall voltage deviation of the distribution area, photovoltaic power output prediction data and load trend data, and outputs a global voltage regulation target. The lower layer is a device-level distributed Actor-Critic algorithm, which generates reactive power output instructions for each photovoltaic and energy storage device based on local voltage characteristics and the upper layer target. The distributed collaborative execution unit is embedded in the photovoltaic inverter and energy storage converter. It exchanges state information with neighboring nodes through a sparse communication protocol, optimizes the objective function based on a consensus protocol, and executes a priority response strategy.

2. The distributed photovoltaic energy storage local autonomous reactive power and voltage control system according to claim 1, characterized in that, The lightweight edge sensing terminal supports multiple communication protocols, including LoRa, HPLC, and RS485, with a communication interval of 5 seconds and a communication distance of 500 meters.

3. The distributed photovoltaic energy storage local autonomous reactive power and voltage control system according to claim 1, characterized in that, In the lightweight edge-sensing terminal, data dimensionality reduction is achieved through sliding window filtering and principal component analysis algorithms.

4. A distributed photovoltaic energy storage local autonomous reactive power and voltage control system according to claim 1, characterized in that, The self-learning dynamic decision-making module supports incremental online learning, embedding pre-trained models into edge devices through transfer learning technology, and updating the policy network weights every 15 minutes based on real-time data.

5. A distributed photovoltaic-storage local autonomous reactive power and voltage control system according to claim 4, characterized in that, The pre-trained model was trained for more than 50,000 steps based on historical transformer area data.

6. A distributed photovoltaic-storage local autonomous reactive power and voltage control system according to claim 1, characterized in that, The objective function for optimizing the consensus protocol is: In the formula, N is the total number of voltage nodes, V i Let V be the voltage at the i-th voltage node. ref I is the reference voltage for the transformer area. i For the circuit flowing through the i-th branch, R i Let be the impedance of the i-th branch; α is the weight of the voltage deviation term, and β is the weight of the line loss term.

7. A distributed photovoltaic-storage local autonomous reactive power and voltage control system according to claim 1, characterized in that, The priority response strategy of the distributed collaborative execution unit includes: when the node voltage exceeds the limit, the adjustment weight of the node is increased by 50%; The reactive power regulation priority of energy storage devices is higher than that of photovoltaic inverters, and they are given priority in providing capacitive or inductive reactive power support.

8. A method for local autonomous reactive power and voltage control of distributed photovoltaic energy storage, characterized in that, The distributed photovoltaic-storage local autonomous reactive power and voltage control system according to any one of claims 1-7 includes the following steps: Step 1: Lightweight Situational Awareness The edge sensing terminal collects key node data at a frequency of 1kHz, removes noise by filtering with a 100ms sliding window, and uses principal component analysis algorithm to compress the data dimension to less than 10 dimensions to extract feature quantities. Step 2: Self-learning decision generation The upper-level deep Q-network at the distribution area level generates a global voltage regulation target based on the overall characteristics of the distribution area. The lower-level Actor-Critic algorithm combines the local voltage status with the global target to calculate the reactive power output command of each device. Step 3: Distributed Cooperative Execution Each photovoltaic and energy storage device exchanges status information every 5 seconds through sparse communication. Based on the consensus protocol, the objective function is optimized, and priority is given to responding to voltage over-limit nodes. The energy storage device and the photovoltaic inverter work together to perform reactive power regulation, thereby achieving dynamic voltage balance in the distribution area.

9. A distributed photovoltaic energy storage local autonomous reactive power and voltage control method according to claim 8, characterized in that, In step 2, incremental online learning is supported. The pre-trained model is embedded into the edge device through transfer learning technology. The pre-trained model is trained for more than 50,000 steps based on historical transformer area data. During online learning, the policy network weights are updated every 15 minutes.

10. A distributed photovoltaic energy storage local autonomous reactive power and voltage control method according to claim 8, characterized in that, In step 3, the communication protocol is compatible with LoRa and HPLC, and the data volume of a single node in a single communication is ≤100 bytes.

Images