Adaptive power coordination control method and system for distributed optical storage access power distribution network

Abstract

This invention discloses an adaptive power coordination control method and system for distributed photovoltaic (PV) and energy storage (ESS) integration into a power distribution network, belonging to the field of power system automation. The method uses a rolling error correction module to dynamically adjust the weights of each branch based on the error mean, variance, and trend, obtaining a corrected predicted value. The corrected value is compared with multi-level fluctuation thresholds, and a mapping relationship is queried to generate and issue energy storage charging and discharging commands. Real-time acquisition of PV and energy storage power data is fed back to the correction module, forming a closed-loop adaptive control system of prediction, decision-making, regulation, feedback, and correction. This invention effectively solves the problems of mismatch between prediction and regulation demand at a single time scale and the lack of error feedback correction leading to accuracy degradation, achieving accuracy and adaptability in PV-ESS coordinated control across multiple time scales.

Description

Technical Field

[0001] This application belongs to the field of power system automation technology, specifically relating to an adaptive power coordination control method and system for distributed photovoltaic energy storage access to distribution networks. Background Technology

[0002] In the power coordination and control of distributed photovoltaic and energy storage connected to the distribution network, the drastic fluctuations in photovoltaic output make accurate power prediction a key prerequisite for the effective coordination and regulation of energy storage systems. How to trigger the adjustment actions of energy storage systems in advance based on the prediction results to smooth out fluctuations is the core challenge in the field.

[0003] Common power coordination control methods typically rely on historical data or simple prediction models at a single time scale for power prediction, and then generate charging and discharging strategies for energy storage systems based on this. However, a single-time-scale prediction framework cannot simultaneously meet the differentiated time accuracy requirements of distribution networks under different operating scenarios, such as ultra-short-term real-time control, short-term scheduling optimization, and medium-term planning. This leads to a time scale mismatch between prediction results and actual control needs. At the same time, existing methods lack a feedback correction mechanism between prediction errors and prediction models. Prediction errors accumulate and gradually amplify over time, causing prediction accuracy to deteriorate. Ultimately, this results in the energy storage system's control actions losing reliable forward-looking basis, making it difficult to achieve prediction-guided precise coordination control. This problem has become a key technical bottleneck restricting the efficient operation of distributed photovoltaic-energy storage systems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an adaptive power coordination control method and system for distributed photovoltaic energy storage access to a distribution network, which can effectively solve the problems mentioned in the background.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides an adaptive power coordination control method for distributed photovoltaic energy storage access to a distribution network, the method comprising the following steps: A multi-timescale power prediction framework is constructed, which consists of three parallel prediction branches: an ultra-short-term power prediction model, a short-term power prediction model, and a medium-term power prediction model. The ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model all adopt a network architecture that cascades and fuses a long short-term memory network and a converter network to generate a fused power prediction result after dynamic fusion through an attention mechanism. The rolling error correction module corrects the fused power prediction result output after dynamic fusion through the attention mechanism. The rolling error correction module calculates three statistical features, namely the mean error, the variance error, and the error change trend, based on the real-time collected prediction error sequence. Based on the mean error, the variance error, and the error change trend, the weight adjustment submodule dynamically adjusts the weight parameters of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model, and outputs the corrected power prediction value. The corrected power prediction value is compared with the power fluctuation threshold. When the power fluctuation percentage exceeds the power fluctuation threshold, a multi-level power fluctuation threshold is configured and a graded judgment is performed to determine the power regulation demand level. Based on the power regulation demand level, query the preset mapping relationship between the predicted power and the energy storage charging and discharging strategy, generate and issue a safe charging and discharging power command to the energy storage system; The actual power data of the distributed photovoltaic system and the charging and discharging power data of the energy storage system are obtained, the net power data of the distribution network is synthesized, and the net power data of the distribution network is fed back to the rolling error correction module to form a closed-loop adaptive control process of prediction, decision-making, regulation, feedback and correction.

[0006] On the other hand, the present invention provides an adaptive power coordination control system for distributed photovoltaic energy storage access to a distribution network, comprising: The multi-timescale power prediction framework consists of three parallel prediction branches: an ultra-short-term power prediction model, a short-term power prediction model, and a medium-term power prediction model. The ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model all adopt a network architecture that cascades and fuses a long short-term memory network and a converter network, which is used to output the fused power prediction result after dynamic fusion through an attention mechanism. The attention mechanism dynamic fusion module, built into the multi-timescale power prediction framework, is used to obtain the current power distribution network operation status information and combine it into a state vector. The state vector is input into the fully connected mapping layer and mapped into a three-dimensional score vector. After processing by the normalized exponential function, the attention weights of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model are obtained. The output results of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model are multiplied by the attention weights corresponding to each prediction branch and then summed to achieve adaptive weighted fusion. The rolling error correction module is used to correct the fused power prediction result output after dynamic fusion through the attention mechanism. The rolling error correction module calculates three statistical features, namely the mean error, the variance error, and the error change trend, based on the real-time collected prediction error sequence, and dynamically adjusts the weight parameters of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model accordingly, and outputs the corrected power prediction value. The threshold judgment module is used to compare the corrected power prediction value with the power fluctuation threshold, and configure multi-level power fluctuation thresholds to perform graded judgment to determine the power regulation demand level; The strategy mapping module is used to query the preset mapping relationship between predicted power and energy storage charging and discharging strategy according to the power adjustment demand level, and generate a safe charging and discharging power command. The instruction issuing module is used to issue the safe charging and discharging power instruction to the energy storage system and receive instruction execution feedback; The central processing unit is used to acquire the actual power data of the distributed photovoltaic system and the charging and discharging power data of the energy storage system, synthesize the net power data of the distribution network, and feed the net power data of the distribution network back to the rolling error correction module to form a closed-loop adaptive control process of prediction, decision-making, regulation, feedback and correction.

[0007] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention constructs a multi-timescale power prediction framework consisting of three parallel prediction branches: ultra-short-term, short-term, and medium-term. It also employs an attention mechanism dynamic fusion module to adaptively adjust the weights of each branch according to the operating status of the distribution network. This allows the distribution network to simultaneously meet the differentiated time accuracy requirements for real-time control, scheduling optimization, and planning under different operating scenarios. It fundamentally solves the mismatch between the single-timescale prediction framework and the multi-timescale control requirements, and significantly improves the applicability and flexibility of power prediction under various operating conditions.

[0008] 2. This invention uses a rolling error correction module to collect the prediction error sequence in real time and calculate three statistical characteristics: the mean error, the variance error, and the error change trend. Based on these characteristics, the weight parameters of each prediction branch are dynamically adjusted, enabling the prediction model to continuously track changes in photovoltaic power generation characteristics and correct its own parameters online in a timely manner. This effectively suppresses the problem of prediction error accumulating and amplifying over time, ensuring the stability and reliability of prediction accuracy across the entire time scale, and providing accurate and forward-looking basis for energy storage regulation.

[0009] 3. This invention determines the power regulation demand level by classifying the corrected power prediction value with multi-level power fluctuation thresholds, and generates safe charging and discharging power commands by querying preset mapping relationships. At the same time, it feeds back the actual power data to the rolling error correction module, constructing a complete closed-loop control process of prediction, decision-making, regulation, feedback and correction. This enables the regulation actions of the energy storage system to respond accurately based on continuously optimized prediction results, realize prediction-oriented adaptive coordinated control, and significantly improve the operating efficiency and stability of distributed photovoltaic energy storage connected to the distribution network. Attached Figure Description

[0010] Figure 1 A schematic diagram of the overall process framework for the adaptive power coordination control method for distributed photovoltaic energy storage access to the distribution network.

[0011] Figure 2 This is a schematic diagram of a multi-timescale power prediction framework based on a fusion network of LSTM and Transformer.

[0012] Figure 3 This is a flowchart illustrating the feedback correction logic of the rolling error correction module.

[0013] Figure 4 The mapping relationship between predicted power and energy storage charging and discharging strategies, and the framework diagram of closed-loop adaptive control flow are presented. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1 To be continued Figure 4 The present invention will be further described in detail below with reference to specific embodiments.

[0015] In a first aspect, this invention provides an adaptive power coordination control method and system for distributed photovoltaic (PV) and energy storage (ESS) integration into a distribution network. This method achieves adaptive power coordination control for distributed PV and ESS integration into the distribution network by constructing a multi-timescale power prediction framework, a rolling error correction module, and a prediction-oriented energy storage coordination control mechanism. The following details each step of the method.

[0016] First, in step S1, a multi-timescale power prediction framework is constructed and fused prediction results are generated, which is carried out according to the following steps.

[0017] Step S101: Construct a multi-timescale power prediction framework containing three independent prediction branches. The multi-timescale power prediction framework consists of three parallel prediction branches: an ultra-short-term power prediction model, a short-term power prediction model, and a medium-term power prediction model.

[0018] The ultra-short-term power forecasting model has a time forecast range of 0 to 15 minutes and is used to capture transient changes in photovoltaic power. The short-term power forecasting model has a time forecast range of 15 minutes to 4 hours and is used to support scheduling optimization over several hours. The medium-term power forecasting model has a time forecast range of 4 to 24 hours and is used to guide day-ahead planning.

[0019] The three prediction branches independently process input data of different granularities, collectively covering prediction needs across the entire time scale from second-level to day-level, providing a data foundation for multi-time-scale regulation of the distribution network. The output of each prediction branch is uniformly a power time series vector corresponding to its respective prediction time range.

[0020] Step S102: Each prediction branch is constructed using a cascaded and fused architecture of a long short-term memory network and a converter network. The ultra-short-term power prediction model, short-term power prediction model, and medium-term power prediction model all employ this cascaded and fused network architecture of long short-term memory networks and converter networks.

[0021] Long Short-Term Memory (LSTM) networks are used to extract temporal features from input data. The core structure of an LTM network includes three gating mechanisms: an input gate, a forget gate, and an output gate, as well as a cell state. The input gate controls the flow of current input information into the cell state.

[0022] The forget gate controls the flow of cell state information from the previous time step that is forgotten. The output gate controls the flow of cell state information from the current time step that is output to the hidden layer. The cell state transmits key temporal features between time steps. In this invention, the Long Short-Term Memory (LSTM) network receives historical power data sequences of different granularities, such as minutes and hours, and extracts the time-series features of photovoltaic power changes.

[0023] The transformer network employs a self-attention mechanism to model the dependencies between different positions in the input sequence. The core components of the transformer network include a multi-head self-attention layer and a feedforward neural network layer. The multi-head self-attention layer captures the dependencies in the input sequence from different perspectives by computing multiple attention heads in parallel.

[0024] Through multi-head parallel computation, the transformer network can simultaneously capture local and global features in the input sequence, enabling effective modeling of long-distance dependencies. In this invention, the transformer network receives temporal features extracted by a long short-term memory network and further mines the correlations between features at different time scales through a self-attention mechanism.

[0025] The fusion of the Long Short-Term Memory (LSTM) network and the transformer network employs a feature-level cascade strategy. Input data is first processed by the LSM network to obtain preliminary extracted temporal feature representations. These temporal feature representations serve as the input sequence for the transformer network, and after processing through the transformer network's multi-head self-attention layer and feedforward neural network layer, an enhanced feature representation incorporating long-range dependencies is obtained. This enhanced feature representation is then decoded through a fully connected mapping in the output layer to obtain power prediction results for each time scale.

[0026] Step S103: Configure the input data, output targets, and rolling update mechanism for each prediction branch. The ultra-short-term power prediction model is based on minute-level historical power data and minute-level meteorological data as input, and uses a rolling prediction mechanism to achieve high-frequency prediction output. The input data of the ultra-short-term power prediction model includes power sampling sequences from the past several minutes and corresponding meteorological monitoring data.

[0027] Meteorological data includes solar irradiance, ambient temperature, wind speed, and humidity parameters. The ultra-short-term power forecasting model operates in a rolling forecasting manner, meaning that within each forecast period, it generates power forecast values ​​for the next few minutes based on the latest measured data, and advances the forecast window by a step size consistent with the sampling interval. The high-frequency forecast results output by the ultra-short-term power forecasting model are used to support real-time dispatching of the distribution network and power balance control at the second to minute level.

[0028] The short-term power forecasting model has a time prediction range of fifteen minutes to four hours. It integrates hourly meteorological forecast data and historical power generation patterns to construct a multi-factor driven forecasting model. The input data for the short-term power forecasting model includes historical power data from the past few hours to one day, meteorological forecast data for the next few hours, and characteristics of historical power generation patterns.

[0029] Meteorological forecast data is provided by meteorological service agencies and includes parameters for solar irradiance forecast, ambient temperature forecast, and cloud cover forecast for the next few hours. Historical power generation pattern characteristics are obtained through pattern recognition of historical power data, including typical photovoltaic power variation curves under different weather conditions and power differences between holidays and weekdays. The forecast results output by the short-term power prediction model are used to support the formulation of short-term dispatch plans for the distribution network and hourly power optimization dispatch.

[0030] The medium-term power prediction model has a time forecast range of four to twenty-four hours. It incorporates numerical weather prediction data and combines seasonal cycle characteristics for trend prediction. The input data for the medium-term power prediction model includes numerical weather prediction data, seasonal cycle characteristics, and historical power data for the same period. The numerical weather prediction data provides forecasts of meteorological elements for the next twenty-four hours, including solar altitude angle, cloud cover, and atmospheric transparency parameters.

[0031] Seasonal cyclical characteristics are derived by analyzing the statistical patterns of photovoltaic power in different seasons and time periods, including seasonal variations such as higher peak power in summer, lower power in winter, and higher power at midday than in the morning and evening. The forecast results output by the medium-term power prediction model are used to support the preparation of day-ahead dispatch plans for the distribution network and the allocation of power generation resources within 24 hours.

[0032] The ultra-short-term, short-term, and medium-term power prediction models were all trained using supervised learning. The training datasets for each model were constructed independently: power and meteorological data for the corresponding time scales were extracted from historical operational databases as input features, and actual measured power values ​​within the same time period were used as the labeling targets. The datasets were then divided into training, validation, and test sets in chronological order. Each model's training objective was to minimize the mean squared error between the predicted and actual values. The training process was iterative until the mean squared error on the validation set stopped decreasing.

[0033] Step S103 completes the data configuration, functional positioning, and training scheme settings for each of the three prediction branches. These three branches serve the control needs of the distribution network at different time scales, forming a multi-level prediction support system from real-time balancing to day-ahead planning. For example, in scenarios with rapid fluctuations in midday irradiance, the ultra-short-term branch can promptly capture the trend of sharp power increases, while the medium-term branch can provide a judgment on the overall power level for the day. The combination of these two can provide a basis for reserving adjustment margins for energy storage systems in advance.

[0034] Step S104: The outputs of the three prediction branches are adaptively weighted and fused through the attention mechanism dynamic fusion module. The attention mechanism dynamic fusion module dynamically weights and fuses the outputs of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model.

[0035] The attention mechanism dynamic fusion module first acquires the current operating status information of the distribution network, including the real-time operating conditions of the distribution network, the current state of charge of the energy storage system, and load forecast information, and combines this information into a state vector. Then, the state vector is input into a fully connected mapping layer, which maps the state vector into a three-dimensional score vector. Each component in the three-dimensional score vector corresponds to the original attention score of the ultra-short-term prediction branch, the short-term prediction branch, and the medium-term prediction branch, respectively. The three-dimensional score vector is then processed by a normalized exponential function to obtain the attention weights for the three prediction branches.

[0036] The attention weights for the three prediction branches are calculated as follows:

[0037] in, Indicates the first The attention weights of the three prediction branches are real numbers between 0 and 1, and the sum of the attention weights of the three branches is equal to 1. is the base of the natural logarithm; Indicates the first The original attention scores for each prediction branch are calculated by the fully connected mapping layer based on the state vector and are real scalars. Indicates the first The original attention scores for each prediction branch. The summation index variable takes values ​​from 1 to 3.

[0038] After obtaining the attention weights of each prediction branch, the output of each prediction branch is multiplied by its corresponding attention weight, and then summed to obtain the fused power prediction result. Let the power prediction values ​​of the three prediction branches at the same future time be... , and Then the predicted power value after fusion at that moment Calculated using the following formula:

[0039] in, Indicates the first The predicted power output of each prediction branch is in megawatts. For the first Attention weights for each prediction branch; The power prediction value is the result of merging, in megawatts. The above calculations are performed independently for each time point within the prediction window, and the calculations at different times do not affect each other.

[0040] When the distribution network is in an emergency control state, the real-time operating condition information of the distribution network, after being processed by the aforementioned fully connected mapping layer and normalized exponential function, gives the ultra-short-term forecast branch a higher attention weight. When the distribution network is in a planned scheduling state, the above mechanism gives the medium-term forecast branch a higher attention weight. The weighted summation operation directly produces the final merged power prediction result.

[0041] The parameters of the fully connected mapping layer in the attention mechanism dynamic fusion module are determined through end-to-end training. The training dataset consists of historical power data, meteorological data, distribution network operating status data, and corresponding actual power values. During training, the mean squared error between the fused prediction result and the actual power value is used as the loss function, calculated as follows:

[0042] in, This represents the total number of time points included in the prediction window, and is a positive integer. This is a time point index variable, with a value range of 1 to... ; Indicates the first The fused predicted power values ​​at each time point are in megawatts. Indicates the first The actual measured power values ​​at each time point, in megawatts; This represents the mean squared error loss value over the entire prediction window, and is a scalar real number in megawatt squared. The training objective is to minimize this loss function. The weight parameters of the fully connected mapping layer are iteratively updated using the backpropagation algorithm until... Training stops when the values ​​on the validation set no longer decrease.

[0043] The fusion power prediction result output after dynamic fusion via the attention mechanism is a time series vector. Its expression is:

[0044] in, The vector represents the fused power prediction results, with dimension 1. ; The total number of time points within the prediction window; Indicates the first Predicted fusion power values ​​at each time point; superscript This represents the transpose of a vector. The length of this vector is consistent with the output length of the ultra-short-term prediction branch, covering a prediction window of 0 to 15 minutes in seconds or minutes.

[0045] Through steps S101 to S104, the multi-timescale power prediction framework processes the historical power data and meteorological data corresponding to the distributed photovoltaic system into a fused power prediction result, providing prediction input for rolling error correction.

[0046] The next step is step S2, which involves performing rolling error correction to obtain the corrected power prediction value. This is carried out according to the following steps.

[0047] Step S201: Configure the overall structure and functional division of the rolling error correction module. The rolling error correction module consists of an error calculation submodule and a weight adjustment submodule. The error calculation submodule collects the deviation between the actual power and the predicted power at each moment within each prediction period, forming a prediction error sequence.

[0048] The weight adjustment submodule dynamically adjusts the weight parameters of the ultra-short-term power prediction model, short-term power prediction model, and medium-term power prediction model based on the statistical characteristics of the prediction error sequence, achieving online adaptive correction of the prediction models. The two submodules form a complete correction link between error acquisition and weight adjustment.

[0049] Step S202: The error calculation submodule collects the actual power and calculates the prediction error. The error calculation submodule performs error acquisition and calculation operations at the end of each prediction cycle. The error calculation submodule first obtains the actual power value at each prediction time within the prediction cycle. The actual power value is collected in real time by the power measurement device of the distribution network.

[0050] The error calculation submodule then calculates the difference between the actual power value and the predicted power value at each prediction time to obtain the prediction error at that prediction time.

[0051] The error calculation submodule adds the prediction error calculated in the current prediction period to a sliding window of a preset length, while removing the oldest prediction error in the window to keep the error sequence within the sliding window updated in real time.

[0052] The length of the sliding window is configured according to the prediction time scale. Shorter sliding windows are used for very short-term predictions to ensure timely corrections. Longer sliding windows are used for medium-term predictions to ensure the stability of corrections. After the sliding window is updated, the error calculation submodule outputs an error sequence containing the most recent prediction errors.

[0053] Step S203: The weight adjustment submodule calculates the statistical characteristics of the error sequence. At the end of each prediction period, the weight adjustment submodule reads the error sequence in the current sliding window from the error calculation submodule and calculates the error sequence to obtain three statistical characteristics: error mean, error variance, and error change trend.

[0054] The mean error is calculated by summing all prediction errors within the sliding window and then dividing by the sliding window length to obtain the average value of the error sequence. The mean error reflects the overall direction and degree of the prediction error shift. When the mean error is greater than zero, it indicates that the predicted values ​​are systematically underestimated. When the mean error is less than zero, it indicates that the predicted values ​​are systematically overestimated. A mean error deviating from zero indicates a systematic bias in the prediction model, requiring correction of the model parameters.

[0055] The error variance is calculated as follows: Calculate the difference between each prediction error and the mean error within the sliding window, square each difference, sum them, and then divide by the sliding window length to obtain the variance of the error sequence. Error variance reflects the degree of dispersion of the prediction error around the mean. A large error variance indicates that the prediction results fluctuate drastically between different times, and the prediction model's response to changes in photovoltaic power is unstable, requiring additional smoothing processing to reduce the fluctuation amplitude.

[0056] The error change trend is calculated as follows: the error sequence within the sliding window is arranged in chronological order, and a first-order linear trend of error change over time is fitted. The slope of this trend is the error change trend.

[0057] Let the length of the sliding window be... The errors at each moment within the window are recorded in chronological order as follows: ,in This represents the error of the earliest time within the window. The error is the latest time value, and the time numbers corresponding to each error are as follows: .

[0058] Error change trend The linear relationship between time sequence number and error value was obtained by fitting the data using the least squares method.

[0059] in, The length of the sliding window is a positive integer. This is the time index variable for the error sequence within the window, with a value range of 1 to... , Represents the earliest time within the window. Representing the latest moment; Indicates the first in the window The prediction error value corresponding to each time number is expressed in megawatts. The slope of the first-order linear trend obtained from the fitting is expressed in megawatts per time step.

[0060] When the value is greater than zero, the prediction error tends to increase over time. When the error is less than zero, the prediction error tends to decrease over time. The larger the absolute value, the faster the error changes.

[0061] When the error change trend When the error shows an increasing trend, the bias of the prediction model is accumulating and amplifying over time, and the model's adaptability is decreasing. It is necessary to increase the update speed of the model parameters to suppress the continuous expansion of the error.

[0062] Step S204: The weight adjustment submodule dynamically adjusts the weight parameters of each prediction model based on statistical characteristics. The weight adjustment submodule comprehensively evaluates three statistical characteristics: mean error, variance error, and error trend. The greater the deviation of the mean error from zero, the more severe the systematic bias of the prediction model, requiring a larger correction to the currently used weight parameters.

[0063] A larger error variance indicates higher uncertainty in the prediction results, necessitating a reduction in the model's weight in the fused output. A continuously increasing error trend indicates deteriorating performance of the prediction model, requiring an acceleration of the weight parameter update rate.

[0064] The weight adjustment submodule calculates the adjustment weights for the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model based on the comprehensive evaluation results of three statistical characteristics. Let the first... The mean error of each prediction branch at the current time is The unit is megawatts; the error variance is The unit is megawatts squared; the error variation trend is... The unit is megawatts per time step; the current original weight is , is obtained by the dynamic fusion module of the attention mechanism in step S104, and is a dimensionless real number with a value range of 0 to 1.

[0065] The weight adjustment submodule calculates the first weight based on three statistical characteristics. Correction coefficients for each prediction branch Correction factor It is a dimensionless positive real number. The correction coefficient is positively correlated with the absolute value of the mean error, the variance of the error, and the positive component of the error trend. The larger the absolute value of the mean error, the larger the variance of the error, and the larger the positive component of the error trend, the larger the correction coefficient will be.

[0066] The weight adjustment submodule adjusts the original weights of each branch. Divide by the corresponding correction factor This yields the unnormalized adjusted weights for each branch. Then, the unnormalized adjusted weights for the three branches are normalized so that the sum of the three adjusted weights equals 1. The normalized result is the [number of branches]. Final adjusted weights for each prediction branch The adjustment weight is dimensionless, ranging from 0 to 1, and the adjustment weights of the three branches satisfy the following conditions: .

[0067] The above calculations are performed independently for the ultra-short-term forecast branch, the short-term forecast branch, and the medium-term forecast branch. The mean error of each branch... Error variance Error change trend Original weights and correction factor Each value is a scalar value that corresponds to each branch independently. Except for the final normalization step, there is no cross-branch vector or matrix operation between branches.

[0068] The weight adjustment submodule will calculate the adjustment weights. The attention weights applied in step S104 are replaced during the prediction result fusion process for the next prediction period. Participate in the fusion calculation. The fused power prediction value in the next prediction period. Calculated using the following formula:

[0069] in, Indicates the first The predicted power output of each prediction branch is in megawatts. For the first The adjustment weights for each prediction branch are dimensionless real numbers. This is the fused power prediction value after rolling error correction, in megawatts. This calculation is performed independently for each time point within the prediction window, and the calculations at different times do not affect each other.

[0070] In summary, through steps S201 to S204, the rolling error correction module receives the fused power prediction results from the multi-timescale power prediction framework, maintains the prediction error sequence in real time using a sliding window mechanism, extracts three statistical features—error mean, error variance, and error change trend—and dynamically adjusts the weight parameters of each prediction branch accordingly, outputting the corrected power prediction value. Through this rolling correction mechanism, the prediction model continuously tracks changes in photovoltaic power generation characteristics and adjusts its parameters in a timely manner, ensuring stable prediction accuracy across all time scales and providing forward-looking data for the power fluctuation threshold judgment in step S3.

[0071] For step S3 in the method of this embodiment, a power fluctuation threshold judgment is performed to determine whether to start energy storage coordination control, and the following steps are followed.

[0072] Step S301: Configure the power fluctuation threshold and set the selection logic for the reference power value. The power fluctuation threshold is the power change amplitude limit set in the distribution network dispatch strategy, used to determine whether the energy storage system needs to be activated for power regulation. The power fluctuation threshold is configured based on the current operating status of the distribution network and the available capacity of the energy storage system, and is dynamically adjusted based on actual operating experience.

[0073] The selection rules for the reference power value are as follows: When the distribution network is in steady-state operation and there is no planned dispatch instruction, the actual power value at the previous moment shall be used as the reference power value.

[0074] When the distribution network is in a planned dispatch period and there is a clearly defined planned power value, the planned power value is used as the baseline power value. When the distribution network is in standby operation mode due to abnormal data acquisition or communication interruption, the historical average power value is used as the baseline power value.

[0075] In step S302, a single-level power fluctuation judgment is performed to determine whether the predicted power fluctuation exceeds the allowable range. The threshold judgment module obtains the reference power value according to the selection rules determined in step S301. The threshold judgment module then calculates the difference between the corrected predicted power value and the reference power value, which reflects the absolute change in the predicted power relative to the reference power.

[0076] The threshold judgment module calculates the ratio of the difference to the reference power value to obtain the power fluctuation percentage. The power fluctuation percentage is a dimensionless real number, representing the relative magnitude of power change as a percentage. The threshold judgment module then compares the power fluctuation percentage with the power fluctuation threshold.

[0077] When the percentage of power fluctuation is less than or equal to the power fluctuation threshold, it is determined that the predicted power fluctuation is within the allowable range, and there is no need to initiate the energy storage coordination control process; monitoring will continue for the next prediction cycle. When the percentage of power fluctuation is greater than the power fluctuation threshold, it is determined that the predicted power fluctuation exceeds the allowable range, and the energy storage coordination control process needs to be initiated, proceeding to step S303 for graded determination.

[0078] Step S303: Configure multi-level power fluctuation thresholds and perform graded judgment to determine the power regulation demand level. The threshold judgment module supports the configuration of multi-level power fluctuation thresholds. Different threshold levels correspond to different power fluctuation severity levels, and different severity levels trigger different energy storage coordination control strategies.

[0079] The multi-level power fluctuation thresholds increase progressively from low to high, comprising at least three levels corresponding to mild, moderate, and severe fluctuations. The values ​​for each level of the multi-level power fluctuation thresholds are pre-calibrated according to the distribution network's dispatching procedures. There is a clear progressive relationship between the calibrated values ​​at each level: lower-level thresholds correspond to smaller power fluctuation amplitudes, triggering milder energy storage regulation measures; higher-level thresholds correspond to larger power fluctuation amplitudes, triggering stronger energy storage regulation measures.

[0080] After step S302 determines that the predicted power fluctuation exceeds the allowable range, the threshold judgment module compares the power fluctuation percentage with the power fluctuation thresholds at each level in sequence. The power fluctuation thresholds at each level are arranged from low to high. The power fluctuation percentage starts from the lowest level threshold. If it exceeds the current level threshold, it continues to be compared with the next higher level threshold until the power fluctuation percentage can no longer exceed a certain level threshold or reaches the highest level threshold.

[0081] When the percentage of power fluctuation exceeds a certain threshold but does not exceed the next higher threshold, the level corresponding to the percentage of power fluctuation is the current power regulation demand level. When the percentage of power fluctuation exceeds the highest level threshold, the power regulation demand level is determined to be the highest level.

[0082] After steps S301 to S303, the threshold judgment module converts the corrected power prediction value into whether the power fluctuation exceeds the limit and the corresponding power adjustment demand level after exceeding the limit, providing a trigger basis and level classification basis for the selection of the energy storage system charging and discharging control strategy in step S4.

[0083] Next, for step S4, the energy storage system charging and discharging control strategy is determined according to the power regulation demand level and instructions are issued. The specific implementation is carried out according to the following steps.

[0084] Step S401: Determine the power regulation demand level based on the power fluctuation amplitude. The strategy mapping module obtains the amplitude value of the current power fluctuation percentage exceeding the power fluctuation threshold from the threshold judgment module. The strategy mapping module determines the power regulation demand level based on this amplitude value. The power regulation demand level includes at least three levels.

[0085] Level 1 power regulation demand corresponds to a power fluctuation amplitude exceeding the threshold but the excess amount is within the preset mild range, triggering a mild charge / discharge strategy. Level 2 power regulation demand corresponds to a power fluctuation amplitude exceeding the threshold but the excess amount is within the preset moderate range, triggering a moderate charge / discharge strategy.

[0086] Level 3 power regulation demand corresponds to a situation where the power fluctuation amplitude significantly exceeds the threshold and the excess amount is within the preset severe range, triggering a severe charge / discharge strategy. Level 1 demand corresponds to lower charge / discharge power, Level 2 demand corresponds to medium charge / discharge power, and Level 3 demand corresponds to higher charge / discharge power.

[0087] The input to this step is the power regulation demand level and the corresponding power fluctuation exceedance value output from step S303. The processing involves determining the level based on a preset range to which the exceedance value belongs, which is pre-calibrated according to the distribution network dispatching regulations. The output is a clearly defined first-level, second-level, or third-level power regulation demand level.

[0088] In step S402, the mapping relationship is queried according to the power regulation demand level and the limiting processing is performed to generate a safe charging and discharging power command. The strategy mapping module queries the preset predicted power and energy storage charging and discharging strategy mapping relationship according to the power regulation demand level determined in step S401 to obtain the corresponding charging and discharging power command and action duration.

[0089] The mapping relationship between predicted power and energy storage charging and discharging strategies includes multi-level power fluctuation thresholds and corresponding energy storage system charging and discharging control strategies for each level. The multi-level power fluctuation thresholds are dynamically configured based on the current operating status of the distribution network and the available capacity of the energy storage system.

[0090] Different power fluctuation ranges correspond to different charging and discharging power levels and operation durations. Charging and discharging power commands include charging power commands and discharging power commands, which correspond to the two operating modes of the energy storage system: absorbing power and releasing power, respectively.

[0091] The strategy mapping module obtains the current state of charge (SOC) and available charge / discharge capacity of the energy storage system. The current SOC reflects the remaining charge level of the energy storage battery. A high SOC limits charging power to prevent overcharging. A low SOC limits discharging power to prevent over-discharging. The available charge / discharge capacity reflects the maximum chargeable and dischargeable capacity of the energy storage battery, and is affected by battery health and temperature.

[0092] The strategy mapping module performs limiting processing on the queried charge / discharge power commands. When the queried charge / discharge power command is a charging power command, the limited charging power command is the smaller of the queried charging power command value and the maximum charging power value allowed by the energy storage system under the current state of charge.

[0093] When the queried charge / discharge power command is a discharge power command, the discharge power command after limiting is the smaller of the queried discharge power command value and the maximum discharge power value allowed by the energy storage system under the current state of charge.

[0094] The inputs for this step are the power regulation demand level output from step S401, the current state of charge (SBC) of the energy storage system, and the available charge / discharge capacity. The processing consists of two stages: the first stage involves querying a preset mapping relationship based on the power regulation demand level to obtain the charge / discharge power command and its duration; the second stage involves limiting the charge / discharge power command based on the SBC and available charge / discharge capacity. The output is a safe charge / discharge power command. The safe charge / discharge power command obtained after limiting is the final output value of this step and is directly used as the content of the command issuance.

[0095] In step S403, the safe charging and discharging power command is sent to the energy storage system and execution feedback is received. The command sending module sends the safe charging and discharging power command generated in step S402 to the energy management unit of the energy storage system. The command sending module and the energy storage system interact with each other through a standard communication interface, which supports commonly used industrial communication protocols.

[0096] The instruction sent by the instruction issuing module includes the charge / discharge power setting value, the action duration, and instruction priority information. The charge / discharge power setting value is the power value in the safe charge / discharge power instruction output in step S402. The action duration indicates the length of time the energy storage system operates continuously according to the charge / discharge power setting value. The instruction priority information is used by the energy storage system to determine the execution order when multiple instructions are received simultaneously.

[0097] After sending a charge / discharge power command to the energy storage system, the command issuance module receives command execution feedback from the energy storage system. This feedback includes confirmation that the command was correctly received and status information indicating whether the energy storage system has begun executing the command. The command issuance module confirms that the command has been correctly received and executed based on the received feedback.

[0098] In summary, after step S4, the strategy mapping module transforms the power regulation demand level determined in step S303 into a safe charging and discharging power command that the energy storage system can actually execute, through mapping query and limiting processing. The command issuance module completes the sending and feedback confirmation of the command, providing the prerequisite for obtaining actual power data and executing closed-loop feedback in step S5.

[0099] Finally, in step S5 of this implementation method, the actual power data is acquired and fed back to the rolling error correction module to form closed-loop control, and is implemented according to the following steps.

[0100] Step S501: The power acquisition device of the distributed photovoltaic system acquires photovoltaic output data in real time, and samples the output power data of the photovoltaic inverter in real time. The power acquisition device transmits the acquired photovoltaic output data to the central processing unit.

[0101] In step S502, the energy management unit of the energy storage system acquires the charging and discharging power data of the energy storage system. The energy management unit monitors the charging and discharging power, state of charge, and temperature operating parameters of the energy storage battery pack in real time. The energy management unit uploads the monitored data to the central processing unit.

[0102] In step S503, the photovoltaic output data and the charging and discharging power data of the energy storage system are combined to obtain the net power data of the distribution network. The central processing unit performs time alignment and synthesis calculations on the photovoltaic output data and the charging and discharging power data of the energy storage system to obtain the net power data of the distribution network. The net power data reflects the actual power balance of the distribution network after the energy storage system participates in power regulation. The central processing unit inputs the net power data to the rolling error correction module.

[0103] Step S504 involves sequentially executing five stages: prediction, decision-making, regulation, feedback, and correction, to complete the closed-loop adaptive control process. The multi-timescale power prediction framework generates a fused power prediction result. The rolling error correction module corrects the prediction result and outputs the corrected power prediction value. The threshold judgment module compares the corrected power prediction value with the power fluctuation threshold.

[0104] When the threshold is exceeded, the strategy mapping module determines the energy storage charging and discharging control strategy, and the command issuing module sends charging and discharging commands to the energy storage system. After the energy storage system executes the charging and discharging commands, the power acquisition device and energy management unit acquire the actual power data and feed it back to the rolling error correction module. The above process constitutes a complete control cycle, which runs continuously in a loop.

[0105] Step S505: Regularly update the training sample set to maintain the long-term adaptability of the prediction model. Periodically collect the latest historical operating data as new training samples. Historical operating data includes power data and meteorological data. Incrementally train the ultra-short-term power prediction model, short-term power prediction model, and medium-term power prediction model based on the newly added training samples. During incremental training, the historical knowledge already learned by the model is retained, while the model parameters are updated to adapt to the latest data distribution characteristics.

[0106] In summary, after step S5, the actual power data is input as a feedback signal to the rolling error correction module, enabling the prediction model to adjust its parameters online based on the actual operating results, forming a complete closed loop of prediction, decision-making, regulation, feedback, and correction. The mechanism of periodically updating the training sample set further ensures the adaptability of the prediction model in long-term operation.

[0107] To facilitate understanding, the technical solution of the present invention will be further explained below with reference to a specific application scenario.

[0108] Suppose a distributed photovoltaic power station is connected to the distribution network. The installed capacity of the photovoltaic power station is several megawatts, and the supporting energy storage system has a capacity of several megawatt-hours.

[0109] On a clear morning, meteorological monitoring shows that solar irradiance gradually increases and ambient temperature gradually rises. According to the method of this invention, the ultra-short-term power prediction model of the multi-timescale power prediction framework, based on power sampling data from the past few minutes and meteorological monitoring data, predicts that photovoltaic power will rise rapidly within the next fifteen minutes. The short-term power prediction model, based on meteorological forecast data and historical power generation patterns, predicts that photovoltaic power will remain at a high level for the next few hours. The medium-term power prediction model, based on numerical weather prediction data and seasonal cycle characteristics, predicts that photovoltaic power will reach its daily peak at midday.

[0110] The attention mechanism dynamic fusion module dynamically calculates the fusion weights of the three prediction branches based on the current operating status of the distribution network and the recent prediction accuracy performance of each prediction branch, and outputs the fused power prediction result. The rolling error correction module, based on the most recent prediction error sequence, determines that the current prediction error is within the normal range and makes only minor weight adjustments. The corrected power prediction value is compared with the power fluctuation threshold. If the predicted photovoltaic power increase is significant and exceeds the power fluctuation threshold, the energy storage coordination control process is triggered.

[0111] The strategy mapping module queries the mapping relationship between predicted power and energy storage charging / discharging strategies based on the power regulation demand level, determining that the energy storage system needs to be charged at the corresponding power to absorb excess photovoltaic power. The command issuing module sends a charging power command to the energy storage system, and the energy storage system begins to execute the charging operation. Photovoltaic output data and energy storage charging / discharging power data are collected in real time and fed back to the rolling error correction module, completing the closed loop of this control cycle.

[0112] Through the above process, predictive-guided adaptive power coordination control for distributed photovoltaic and energy storage access to the distribution network is achieved.

[0113] On the other hand, the adaptive power coordination control system for distributed photovoltaic energy storage access to the distribution network disclosed in this invention includes: The multi-timescale power prediction framework consists of three parallel prediction branches: an ultra-short-term power prediction model, a short-term power prediction model, and a medium-term power prediction model. The ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model all adopt a network architecture that cascades and fuses a long short-term memory network and a converter network, which is used to output the fused power prediction result after dynamic fusion through an attention mechanism. The attention mechanism dynamic fusion module, built into the multi-timescale power prediction framework, is used to obtain the current power distribution network operation status information and combine it into a state vector. The state vector is input into the fully connected mapping layer and mapped into a three-dimensional score vector. After processing by the normalized exponential function, the attention weights of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model are obtained. The output results of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model are multiplied by the attention weights corresponding to each prediction branch and then summed to achieve adaptive weighted fusion. The rolling error correction module is used to correct the fused power prediction results output after dynamic fusion through the attention mechanism. The rolling error correction module calculates three statistical features, namely the mean error, the variance error, and the error change trend, based on the real-time collected prediction error sequence. Based on these features, it dynamically adjusts the weight parameters of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model, and outputs the corrected power prediction value. The threshold judgment module is used to compare the corrected power prediction value with the power fluctuation threshold, and configure multi-level power fluctuation thresholds to perform graded judgment to determine the power regulation demand level; The strategy mapping module is used to query the preset mapping relationship between predicted power and energy storage charging and discharging strategy according to the power adjustment demand level, and generate safe charging and discharging power instructions. The instruction issuing module is used to issue safe charging and discharging power instructions to the energy storage system and receive instruction execution feedback; The central processing unit is used to acquire the actual power data of the distributed photovoltaic system and the charging and discharging power data of the energy storage system, synthesize the net power data of the distribution network, and feed the net power data of the distribution network back to the rolling error correction module to form a closed-loop adaptive control process of prediction, decision-making, regulation, feedback and correction.

[0114] 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 present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0115] 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. An adaptive power coordination control method for distributed photovoltaic energy storage connected to a distribution network, characterized in that, Includes the following steps: A multi-timescale power prediction framework is constructed, which consists of three parallel prediction branches: an ultra-short-term power prediction model, a short-term power prediction model, and a medium-term power prediction model. The ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model all adopt a network architecture that cascades and fuses a long short-term memory network and a converter network to generate a fused power prediction result after dynamic fusion through an attention mechanism. The rolling error correction module corrects the fused power prediction result output after dynamic fusion through the attention mechanism. The rolling error correction module calculates three statistical features, namely the mean error, the variance error, and the error change trend, based on the real-time collected prediction error sequence. Based on the mean error, the variance error, and the error change trend, the weight adjustment submodule dynamically adjusts the weight parameters of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model, and outputs the corrected power prediction value. The corrected power prediction value is compared with the power fluctuation threshold. When the power fluctuation percentage exceeds the power fluctuation threshold, a multi-level power fluctuation threshold is configured and a graded judgment is performed to determine the power regulation demand level. Based on the power regulation demand level, query the preset mapping relationship between the predicted power and the energy storage charging and discharging strategy, generate and issue a safe charging and discharging power command to the energy storage system; The actual power data of the distributed photovoltaic system and the charging and discharging power data of the energy storage system are obtained, the net power data of the distribution network is synthesized, and the net power data of the distribution network is fed back to the rolling error correction module to form a closed-loop adaptive control process of prediction, decision-making, regulation, feedback and correction.

2. The adaptive power coordination control method for distributed photovoltaic energy storage access to a distribution network according to claim 1, characterized in that, The ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model all adopt a network architecture that cascades and integrates a long short-term memory network and a converter network, specifically: The input data is first processed by the Long Short-Term Memory network to obtain a preliminary extracted temporal feature representation; The initially extracted temporal feature representation is used as the input sequence of the transformer network. After being processed by the multi-head self-attention layer and feedforward neural network layer of the transformer network, an enhanced feature representation that incorporates long-distance dependencies is obtained. The enhanced feature representation, which incorporates long-range dependencies, is then decoded into power prediction results for each time scale through a fully connected mapping of the output layer.

3. The adaptive power coordination control method for distributed photovoltaic energy storage access to a distribution network according to claim 1, characterized in that, The generation of the fusion power prediction result after dynamic fusion via the attention mechanism specifically includes: The current power distribution network operation status information is obtained through the attention mechanism dynamic fusion module, and the power distribution network operation status information is combined into a state vector. The power distribution network operation status information includes the real-time operation status of the power distribution network, the current state of charge of the energy storage system, and load forecast information. The state vector is input into a fully connected mapping layer and mapped into a three-dimensional score vector. After being processed by a normalized exponential function, the attention weights of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model are obtained. The outputs of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model are multiplied by the attention weights corresponding to the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model, respectively, and then summed to obtain the fused power prediction result output after dynamic fusion by the attention mechanism. Specifically, when the distribution network is in an emergency control state, the attention mechanism dynamic fusion module gives the ultra-short-term prediction branch a higher attention weight; when the distribution network is in a planned scheduling state, the attention mechanism dynamic fusion module gives the medium-term prediction branch a higher attention weight.

4. The adaptive power coordination control method for distributed photovoltaic energy storage access to a distribution network according to claim 1, characterized in that, The rolling error correction module outputs the corrected power prediction value, specifically including: At the end of each prediction period, the error calculation submodule collects the actual power value and calculates the prediction error. It adds the prediction error calculated for the current prediction period to a sliding window of a preset length and removes the oldest prediction error from the sliding window to keep the error sequence within the sliding window updated in real time. The length of the sliding window is configured according to the prediction time scale. Shorter sliding windows are used for ultra-short-term predictions and longer sliding windows are used for medium-term predictions. The weight adjustment submodule reads the error sequence within the current sliding window and calculates the mean error, the variance error, and the error change trend of the error sequence. The error change trend is obtained by fitting the linear relationship between the time sequence number and the error value using the least squares method. The weight adjustment submodule calculates the correction coefficients for the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model based on the mean error, the variance error, and the positive portion of the error change trend. The original weights of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model are divided by the correction coefficients calculated for each prediction branch and then normalized to obtain the adjusted weights corresponding to the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model, respectively. In the next prediction period, the outputs of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model are multiplied by the corresponding adjustment weights of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model, respectively, and then summed to obtain the corrected power prediction value.

5. The adaptive power coordination control method for distributed photovoltaic energy storage access to a distribution network according to claim 1, characterized in that, The configuration of multi-level power fluctuation thresholds and the execution of graded judgment to determine the power regulation demand level specifically include: The multi-level power fluctuation threshold increases progressively from low to high, and includes at least three levels, corresponding to three severity levels: mild fluctuation, moderate fluctuation, and severe fluctuation. The power fluctuation percentage is compared sequentially with the power fluctuation thresholds at each level. When the power fluctuation percentage exceeds a certain level threshold but does not exceed the next higher level threshold, the level corresponding to the power fluctuation percentage is the current power adjustment demand level. When the power fluctuation percentage exceeds the highest level threshold, the power adjustment demand level is determined to be the highest level.

6. The adaptive power coordination control method for distributed photovoltaic energy storage access to a distribution network according to claim 5, characterized in that, The step of querying the preset mapping relationship between predicted power and energy storage charging and discharging strategy based on the power regulation demand level, and generating and issuing safe charging and discharging power commands to the energy storage system specifically includes: The power regulation demand levels include at least Level 1, Level 2, and Level 3, corresponding to light, moderate, and heavy charging and discharging strategies, respectively; the mapping relationship between predicted power and energy storage charging and discharging strategies includes charging and discharging power commands and action durations corresponding to each level; The strategy mapping module queries the mapping relationship between the predicted power and the energy storage charging and discharging strategy based on the determined power regulation demand level, obtains the corresponding charging and discharging power command and action duration, and acquires the current state of charge and available charging and discharging capacity of the energy storage system. It then performs amplitude limiting processing on the queried charging and discharging power command and takes the smaller of the queried charging and discharging power command value and the maximum charging and discharging power value allowed by the energy storage system under the current state of charge as the safe charging and discharging power command.

7. The adaptive power coordination control method for distributed photovoltaic energy storage access to a distribution network according to claim 1, characterized in that, The power fluctuation threshold is dynamically configured based on the current operating status of the distribution network and the available capacity of the energy storage system, and the selection rules for the reference power value are as follows: When the power distribution network is in steady-state operation and there is no planned dispatch instruction, the actual power value at the previous moment is used as the reference power value. When the power distribution network is in a planned scheduling period and there is a clear scheduled power value, the scheduled power value is used as the benchmark power value. When the power distribution network is in standby operation mode due to abnormal data acquisition or communication interruption, the historical average power value is used as the reference power value.

8. The adaptive power coordination control method for distributed photovoltaic energy storage access to a distribution network according to claim 1, characterized in that, The ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model are configured as follows: The ultra-short-term power prediction model is based on minute-level historical power data and minute-level meteorological data, and uses a rolling prediction mechanism to achieve high-frequency prediction output. The time prediction range of the ultra-short-term power prediction model is zero to fifteen minutes. The short-term power prediction model integrates hourly meteorological forecast data and historical power generation patterns to construct a multi-factor driven prediction model. The time prediction range of the short-term power prediction model is from fifteen minutes to four hours. The medium-term power prediction model incorporates numerical weather forecast data and combines seasonal cycle characteristics to make trend predictions. The time prediction range of the medium-term power prediction model is four to twenty-four hours.

9. The adaptive power coordination control method for distributed photovoltaic energy storage access to a distribution network according to claim 4, characterized in that, The length of the sliding window is configured according to the prediction time scale, specifically: The sliding window length corresponding to the ultra-short-term power prediction model is smaller than that corresponding to the short-term power prediction model, and the sliding window length corresponding to the short-term power prediction model is smaller than that corresponding to the medium-term power prediction model, so as to achieve a balance between the timeliness and stability of the correction.

10. An adaptive power coordination control system for distributed photovoltaic energy storage connected to a distribution network, characterized in that, The system for performing the method according to any one of claims 1 to 9, the system comprising: The multi-timescale power prediction framework consists of three parallel prediction branches: an ultra-short-term power prediction model, a short-term power prediction model, and a medium-term power prediction model. The ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model all adopt a network architecture that cascades and fuses a long short-term memory network and a converter network, which is used to output the fused power prediction result after dynamic fusion through an attention mechanism. The attention mechanism dynamic fusion module, built into the multi-timescale power prediction framework, is used to obtain the current power distribution network operation status information and combine it into a state vector. The state vector is input into the fully connected mapping layer and mapped into a three-dimensional score vector. After processing by the normalized exponential function, the attention weights of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model are obtained. The output results of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model are multiplied by the attention weights corresponding to each prediction branch and then summed to achieve adaptive weighted fusion. The rolling error correction module is used to correct the fused power prediction result output after dynamic fusion through the attention mechanism. The rolling error correction module calculates three statistical features, namely the mean error, the variance error, and the error change trend, based on the real-time collected prediction error sequence, and dynamically adjusts the weight parameters of the ultra-short-term power prediction model, the short-term power prediction model, and the medium-term power prediction model accordingly, and outputs the corrected power prediction value. The threshold judgment module is used to compare the corrected power prediction value with the power fluctuation threshold, and configure multi-level power fluctuation thresholds to perform graded judgment to determine the power regulation demand level; The strategy mapping module is used to query the preset mapping relationship between predicted power and energy storage charging and discharging strategy according to the power adjustment demand level, and generate a safe charging and discharging power command. The instruction issuing module is used to issue the safe charging and discharging power instruction to the energy storage system and receive instruction execution feedback; The central processing unit is used to acquire the actual power data of the distributed photovoltaic system and the charging and discharging power data of the energy storage system, synthesize the net power data of the distribution network, and feed the net power data of the distribution network back to the rolling error correction module to form a closed-loop adaptive control process of prediction, decision-making, regulation, feedback and correction.

Images