A power dispatching-based optical-electrical transmission performance evaluation method and system

Abstract

The application discloses a kind of photovoltaic transmission performance evaluation method and system based on power dispatching, it is related to photovoltaic power plant operation and maintenance evaluation technical field, the present application synchronously collects photovoltaic power station operation data, power grid dispatching instruction data and station meteorological data and completes time alignment, extracts active power transmission efficiency, reactive response time and other transmission performance characteristics, constructs the performance characteristic vector sequence with time scale;Analysis identifies power limiting, peak shaving, reactive power regulation type dispatching event and generates label sequence, with meteorological parameter as confusion variable, the independent contribution degree of scheduling event to performance characteristics is calculated by causal inference algorithm, and the purified equipment degradation characteristic sequence is obtained by stripping scheduling fluctuation component;Based on purification sequence, generate degradation index and extrapolate trend through time series prediction model, combined with multi-level threshold trigger alarm and output evaluation report.The present application can effectively distinguish scheduling disturbance and equipment real degradation, improve evaluation accuracy, reduce false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant operation and maintenance assessment technology, specifically to a method and system for assessing the performance of photovoltaic transmission based on power dispatch. Background Technology

[0002] The photovoltaic transmission performance of a photovoltaic power plant, including active power transmission efficiency, reactive power response capability, voltage stability, and harmonic levels, directly affects the power plant's generation revenue and grid connection safety. Current standards and engineering practices typically employ online monitoring schemes to assess transmission performance: power quality monitoring devices and power meters are deployed at the grid connection point and key nodes to continuously collect electrical parameters and calculate indicators such as transmission efficiency, line loss rate, voltage deviation rate, and harmonic distortion rate, thereby determining the health status and performance degradation degree of the transmission system.

[0003] However, under the background of the construction of new power systems, photovoltaic power plants are frequently affected by various instructions issued by the power grid dispatch center, including power curtailment instructions, peak shaving instructions, and reactive power and voltage regulation instructions. When dispatch instructions change the operating conditions of the power plant, transmission performance indicators will inevitably fluctuate accordingly. During power curtailment, the inverter deviates from its maximum power point, leading to a decrease in efficiency. During reactive power regulation, the voltage and harmonics at the grid connection point change accordingly. This fluctuation caused by dispatch is a normal response of the power plant to grid regulation, rather than a true deterioration in the health of the equipment.

[0004] Existing technologies, such as patent applications with publication numbers CN122243183A and CN122068584A related to photovoltaic power plant operation and maintenance assessment, do not distinguish between scheduling factors and equipment degradation in their current online monitoring methods. This leads to two typical types of misjudgments: false alarms: incorrectly classifying efficiency decline caused by scheduling power curtailment as equipment failure; and false negatives: true early degradation is submerged in normal fluctuations caused by scheduling and thus fails to be detected. Some improved solutions use a simple rule of directly removing data during power curtailment periods, but the time windows for power curtailment and degradation may completely overlap, and the superposition of scheduling effects and degradation effects cannot be handled. Summary of the Invention

[0005] In view of the above-mentioned technical deficiencies, the purpose of this invention is to provide a method and system for evaluating the performance of photoelectric transmission based on power dispatch.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The first aspect of the present invention provides a method for evaluating the performance of photovoltaic transmission based on power dispatch, including: S1: synchronously collecting and time-aligning real-time operation data of photovoltaic power plants, historical dispatch command data issued by the power grid dispatch center, and environmental meteorological data of the power plant, and generating a multi-source fusion dataset.

[0007] S2: Based on the multi-source fusion dataset, extract transmission performance features including active power transmission efficiency, reactive power response time, voltage deviation rate and total harmonic distortion rate of current, and construct a time-stamped performance feature vector sequence.

[0008] S3: Parse historical dispatch instruction data, identify and mark dispatch events of the types of power rationing instructions, peak shaving instructions and reactive power regulation instructions, and generate a dispatch event tag sequence.

[0009] S4: Correlate the performance feature vector sequence with the scheduling event label sequence over time. For each type of scheduling event, use a causal inference algorithm to calculate the independent contribution of each scheduling event to each transmission performance feature, with at least irradiance and temperature from the station's environmental meteorological data as confounding variables.

[0010] S5: Based on the independent contribution, remove the fluctuation component caused by the scheduling event from the performance feature vector sequence to obtain the purified device's own transmission performance degradation feature sequence.

[0011] S6: Based on the purified degradation feature sequence, the transmission performance degradation index at the current moment is generated using a time-series prediction model, and the future degradation trend curve is extrapolated.

[0012] S7: Compare the transmission performance degradation index with preset multi-level thresholds. When the warning threshold is exceeded, generate an alarm and output a performance evaluation report containing a decoupling explanation of the scheduling event.

[0013] A second aspect of the present invention provides a system for performing the power dispatch-based photovoltaic transmission performance evaluation method described in the present invention, comprising: a data acquisition module for synchronously acquiring and time-aligning real-time operating data of photovoltaic power plants, historical dispatch command data issued by the power grid dispatch center, and environmental meteorological data of the power plant, and generating a multi-source fusion dataset.

[0014] The feature extraction module is used to extract transmission performance features, including active power transmission efficiency, reactive power response time, voltage deviation rate, and total harmonic distortion rate of current, based on a multi-source fusion dataset, and construct a time-stamped performance feature vector sequence.

[0015] The scheduling event parsing module is used to parse historical scheduling instruction data, identify and mark scheduling events of the types of power restriction instructions, peak shaving instructions and reactive power regulation instructions, and generate a scheduling event tag sequence.

[0016] The causal decoupling calculation module is used to temporally correlate the performance feature vector sequence with the scheduling event label sequence. For each type of scheduling event, the causal inference algorithm is used to calculate the independent contribution of each scheduling event to each transmission performance feature. Based on the independent contribution, the fluctuation component caused by the scheduling event is removed from the performance feature vector sequence to obtain the purified equipment transmission performance degradation feature sequence, wherein at least the irradiance and temperature in the field environmental meteorological data are used as confounding variables.

[0017] The degradation prediction module is used to generate the transmission performance degradation index at the current moment and extrapolate the future degradation trend curve based on the purified degradation feature sequence using a time-series prediction model.

[0018] The evaluation and alarm module compares the transmission performance degradation index with preset multi-level thresholds. When the warning threshold is exceeded, an alarm is generated, and a performance evaluation report containing a decoupled explanation of the scheduling event is output.

[0019] The beneficial effects of the present invention are as follows: (1) The present invention standardizes and integrates power grid dispatch command data with photovoltaic power plant operation data and meteorological data on a unified time grid, and constructs a multi-source fusion dataset that simultaneously carries electrical operation information, dispatch control information and meteorological environment information, so that subsequent performance evaluation can explicitly distinguish the different driving sources of external dispatch intervention and internal equipment degradation, laying a complete input data foundation for causal separation.

[0020] (2) This invention uses a dual machine learning method based on a counterfactual reasoning framework, with irradiance and temperature as confounding variables, to quantify the independent causal contribution of scheduling events such as power rationing, peak shaving and reactive power regulation to each transmission performance characteristic. This solves the problem that traditional correlation analysis methods cannot eliminate meteorological confounding bias, and achieves precise decoupling of reasonable operating condition fluctuations caused by scheduling from the actual performance degradation of equipment at the feature level, thereby reducing false alarm rate and false alarm rate.

[0021] (3) This invention obtains a purified degradation feature sequence by removing the contribution of scheduling events frame by frame from the original performance features, so that even during the scheduling power rationing event, the actual degradation status of the equipment can be continuously monitored. This overcomes the defects of the traditional method of removing data during power rationing period, which leads to the degradation detection blind spot, and the limitation of the rule method in being unable to handle the superposition effect of scheduling and degradation.

[0022] (4) This invention uses a weighted fusion degradation index based on purification degradation characteristics and LSTM time series prediction to make degradation trend prediction unaffected by frequent changes in power grid dispatching strategies. It triggers hierarchical alarms through multi-level thresholds and outputs a performance evaluation report containing dispatching decoupling explanations, providing maintenance personnel with traceable and interpretable decision-making basis.

[0023] (5) In a further optimization scheme of the present invention, the parameters of the causal inference and degradation prediction model are continuously tested and updated using the data of the degradation stabilization period under normal conditions through a closed-loop adaptive optimization mechanism, so that the evaluation system can maintain its adaptability to changing operating conditions and equipment aging conditions throughout the entire life cycle of the power plant. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention.

[0026] Figure 2 This is a schematic diagram of the system structure connection of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Reference Figure 1 As shown, the first aspect of the present invention provides a method for evaluating the performance of photovoltaic transmission based on power dispatch, including: S1: synchronously collecting and time-aligning real-time operation data of photovoltaic power plants, historical dispatch command data issued by the power grid dispatch center, and environmental meteorological data of the power plant to generate a multi-source fusion dataset.

[0029] In a specific embodiment of the present invention, in step S1: the real-time operating data includes the DC side voltage and current of the photovoltaic array, the AC side output power of the inverter, the voltage and current of each node of the collector line, and the active power and reactive power at the grid connection point.

[0030] The historical dispatch instruction data includes power rationing instructions issued by AGC and their execution time windows, reactive power and voltage regulation instructions issued by AVC and their execution time windows, as well as planned output curves.

[0031] The environmental meteorological data of the site includes total horizontal irradiance, component backplane temperature, ambient temperature, and wind speed.

[0032] Specifically, the real-time operation data of the photovoltaic power station is collected from the power station monitoring system and includes four data items.

[0033] The DC-side voltage and current of the photovoltaic array are obtained through the communication interface of the intelligent combiner box or string inverter. The AC-side output power of the inverter is obtained through the Modbus or IEC 61850 protocol. The voltage and current of each node of the collector line are obtained through the line protection and control device. The active and reactive power at the grid connection point are obtained through the integrated automation system of the substation. Historical dispatch instructions issued by the power grid dispatch center are obtained from the dispatch automation system through the dispatch data network. This includes three types of dispatch information: the first type is the power curtailment instructions issued by AGC (Automatic Generation Control), which mainly manages the active power of the power station. These instructions carry the target active power, the issuance timestamp, and the execution time window; the second type is the reactive power and voltage regulation instructions issued by AVC (Automatic Voltage Control), which is used to regulate the reactive power of the power station and the grid connection voltage. These instructions include the regulation target value and the execution time window; the third type is the planned output curve of the photovoltaic power station. The environmental meteorological data of the site is collected from the meteorological station of the photovoltaic power station, including: total horizontal irradiance, measured in watts per square meter; module backsheet temperature, measured in degrees Celsius; ambient temperature, measured in degrees Celsius; and wind speed, measured in meters per second.

[0034] The three types of data mentioned above have different original sampling periods, necessitating the establishment of a unified time reference. The unified time reference is determined by starting from the earliest common timestamp and using a fixed step size. Equal-interval time series. The determination was based on the following: Through parameter comparison and optimization experiments, the root mean square error between the independent contribution of the causal inference estimate in subsequent step S4 and the true value of the power-off isolation test was compared among candidate step sizes of 1 second, 5 seconds, 10 seconds, 30 seconds, 60 seconds, and 300 seconds. The results showed that the error decreased to the lowest level and entered the plateau region when the step size was 30 seconds. Therefore, [the value was determined]. Take approximately 30 seconds. For operational and meteorological data, if the original sampling period is less than... If the average of multiple sampling points within the window is taken as the downsampled value, then if it is greater than 1, the average of the sampling points within the window is taken as the downsampled value. If the time window information is insufficient, linear interpolation is used to fill in the gaps. For scheduling instruction data, the original time window information is preserved without over-resampling.

[0035] To eliminate dimensional differences, Z-score standardization was performed on the three types of data at the channel dimension. The mean and standard deviation used were obtained by statistically analyzing historical data collected from multiple photovoltaic power plants under stable clear-sky conditions with no faults and no power curtailment. After standardization, the data from each category were concatenated at a unified time to generate a multi-source fused dataset.

[0036] It should be noted that the above The determination process is completed offline before system deployment, using historical archived data to iterate through each candidate step size and compare the purification effects to determine the optimal value; during the online operation phase, this fixed value is used directly. Data alignment can be performed without repeating the step size optimization process.

[0037] For example, in a 50 MW photovoltaic power plant, there are approximately 200 channels for operational data, 2 channels for dispatch command data, and 4 channels for meteorological data. Data was collected continuously for 48 hours at 30-second intervals, resulting in approximately 5760 time frames. Each time frame contains a joint feature vector of approximately 206 dimensions. This dataset, when input into subsequent processes, successfully supported the estimation of the independent contribution of dispatch events in step S4.

[0038] S2: Based on the multi-source fusion dataset, extract transmission performance features including active power transmission efficiency, reactive power response time, voltage deviation rate and total harmonic distortion rate of current, and construct a time-stamped performance feature vector sequence.

[0039] In a specific embodiment of the present invention, in step S2: the active power transmission efficiency is the ratio of the active power output at the grid connection point to the theoretical maximum DC power of the photovoltaic array calculated based on real-time irradiance and component temperature.

[0040] The reactive power response time is the time elapsed from the moment the reactive power adjustment command is issued until the inverter's reactive power output first reaches a preset percentage of the command target value, such as 90%.

[0041] The voltage deviation rate is the absolute percentage of the effective value of the grid connection voltage deviating from the rated voltage.

[0042] The total harmonic distortion rate of the current is the ratio of the square root of the sum of the squares of the effective values ​​of each harmonic component in the grid-connected current to the effective value of the fundamental component.

[0043] Specifically, starting from the multi-source fusion dataset obtained in step S1, the standardized operational data is first restored to its original physical dimensions. This embodiment extracts four types of transmission performance features.

[0044] The first type of characteristic is active power transmission efficiency. This is defined as the ratio of the active power output at the grid connection point to the theoretical maximum DC power of the photovoltaic array. The formula for calculating (unit: kilowatt) is:

[0045] In the formula The rated DC power of the photovoltaic array under standard test conditions is expressed in kilowatts and is obtained by summing the parameters on the module nameplate. G is the total horizontal irradiance, expressed in watts per square meter. For reference irradiance, 1000 watts per square meter is used, representing the value specified under standard test conditions. γ is the module power temperature coefficient, expressed per degree Celsius; a typical crystalline silicon module has a power temperature coefficient of approximately -0.0035 per degree Celsius, provided in the module manufacturer's datasheet. The temperature of the module backsheet is in degrees Celsius. The reference temperature is 25 degrees Celsius. That is That is, the actual measured active power at the grid connection point, in kilowatts and The ratio of is dimensionless. Under normal operating conditions. Typical values ​​are around 0.85 to 0.95, and their downward trend reflects inverter efficiency degradation, line joint oxidation, or transformer malfunction.

[0046] The second characteristic is reactive response time. The unit is seconds, defined as the time elapsed from the moment the AVC reactive power regulation command is issued until the inverter's reactive power output first reaches 90% of the command target value. The specific operation involves reading the command issuance timestamp. and target reactive power Extract the inverter's reactive power output after that moment from the operating data. The time series is searched point by point backwards for the first condition that satisfies the condition. The moment ,but If the target is not reached within the maximum search time of 30 seconds, a timeout will be marked.

[0047] The third type of characteristic is voltage deviation rate. It is defined as the absolute percentage of the effective value of the grid connection point voltage deviating from the rated voltage, which is determined based on the grid connection voltage level.

[0048] The fourth characteristic is the total harmonic distortion of the current. According to GB / T 14549, it is defined as the ratio of the square root of the sum of the squares of the effective values ​​of each harmonic current to the effective value of the fundamental frequency, with the highest harmonic order being 25.

[0049] The four types of features at each unified time step are combined into a performance feature vector, and then arranged in chronological order to construct a time-stamped sequence of performance feature vectors.

[0050] For example, in a 50 MW photovoltaic power plant instance, The power output is 50 MW, γ is -0.0035 degrees Celsius, and the grid connection point has a rated voltage of 35 kV. Four types of characteristics were calculated for each of the 5760 time frames. During the analysis period, The average value was 0.89, showing a significant decrease in 4 out of 7 power curtailment events, reaching a low of 0.72. However, maintenance records confirmed that the equipment itself was not malfunctioning, which is precisely the dispatch effect that needs to be stripped away in step S4. The reactive power response time during normal operation was approximately 1.2 to 2.8 seconds.

[0051] S3: Parse historical dispatch instruction data, identify and mark dispatch events of the types of power rationing instructions, peak shaving instructions and reactive power regulation instructions, and generate a dispatch event tag sequence.

[0052] In a specific embodiment of the present invention, the parsing of historical dispatch instruction data in step S3 specifically includes: parsing the instruction type field in the dispatch instruction message to distinguish between power rationing instructions, peak shaving instructions, and reactive power regulation instructions.

[0053] Extract the start and end times of the execution time window corresponding to each instruction.

[0054] Multiple scheduling events with overlapping time windows are merged, and different types of scheduling events with overlapping time windows are retained and marked with overlap to generate a sequence of scheduling event labels.

[0055] Specifically, structured parsing is performed on the historical dispatch command messages synchronously collected in step S1. Grid dispatch messages typically follow IEC 60870-5-101 or 104 remote control communication protocols, and their type identifier field is used to distinguish dispatch commands with different functions. The parsing process is as follows: extract the message type identifier field. If the instruction is an AGC setting command and the target active power is less than approximately 80% of the current available output of the photovoltaic power station, it is marked as a power curtailment command type, and the current available output is taken as the 98th percentile of the active power of the grid-connected point under similar irradiance conditions during the non-dispatch period within 30 minutes before the command was issued; if it is an AGC setting command but the target is not less than 80% of the available output and it is during morning or evening peak hours, it is marked as a peak shaving command type; if it is an AVC voltage or reactive power setting command, it is marked as a reactive power regulation command type.

[0056] Extract the start time (instruction issuance timestamp) and end time (next new instruction of the same type) of each instruction execution time window. If the window exceeds the maximum window of 15 minutes, it is truncated. This window references the typical cycle of AGC / AVC instructions in DL / T 1700 and DL / T 1867. Multiple scheduling events with overlapping time windows are merged: overlapping events of the same type have their time windows merged; overlapping events of different types are retained separately and marked with an overlap flag for interaction processing in step S4's causal inference.

[0057] For example, in a 48-hour instance of a 50 MW photovoltaic power plant, 7 power curtailment events, 12 peak shaving events, and 6 reactive power regulation events were parsed from approximately 380 dispatch messages. Including one instance where power curtailment and reactive power regulation overlapped for approximately 8 minutes, a dispatch event tag sequence of 25 records was generated.

[0058] It should be noted that if there are no similar irradiance levels within 30 minutes prior to the issuance of the instruction, i.e., the irradiance deviation does not exceed ±50W / m², then the instruction will be valid. 2 During periods without scheduling, 98% of the theoretical maximum DC power calculated by substituting real-time irradiance and component temperature into the photovoltaic array theoretical power model is used as the estimate of the current available output.

[0059] S4: Correlate the performance feature vector sequence with the scheduling event label sequence over time. For each type of scheduling event, use a causal inference algorithm to calculate the independent contribution of each scheduling event to each transmission performance feature, with at least irradiance and temperature from the station's environmental meteorological data as confounding variables.

[0060] In a specific embodiment of the present invention, in step S4, the causal inference algorithm is a causal effect estimation algorithm based on a counterfactual reasoning framework. It takes each transmission performance feature in the performance feature vector sequence as the result variable, the scheduling event as the processing variable, and at least the irradiance and temperature in the station environmental meteorological data as confounding variables. It estimates the average processing effect of each scheduling event on each transmission performance feature and takes the average processing effect as the independent contribution.

[0061] In a specific embodiment of the present invention, the causal effect estimation algorithm based on the counterfactual reasoning framework is a dual machine learning method. This method estimates the independent contribution by fitting a first machine learning model with a confounding variable as input and a transmission performance feature as output, and calculating a first residual, which is the difference between the observed value of the transmission performance feature and the predicted value of the first machine learning model.

[0062] The second machine learning model is fitted with the confounding variable as input and the binary indicator variable of the scheduling event as output, and the second residual is calculated as the difference between the observed value of the scheduling event indicator variable and the predicted value of the second machine learning model.

[0063] Regressing the first residual against the second residual yields regression coefficients that represent the independent contributions after eliminating confounding bias.

[0064] In a specific embodiment of the present invention, the confounding variables also include the number of inverters in operation and the degree of dust accumulation in the photovoltaic array. The degree of dust accumulation is obtained by comparing the difference in DC-side output power of the array under different cleaning cycles under the same irradiance conditions.

[0065] Specifically, three core variables are first constructed. The result variable Y takes a certain transmission performance characteristic from step S2. , , , Each variable is modeled independently; the processing variable D is a binary indicator variable of a certain type of scheduling event in step S3, and the variable value rule is that it is marked as 1 when it is within the scheduling event time window and marked as 0 at other times; the confusion variable X is irradiance, component backplane temperature and ambient temperature, and all data are restored to the original physical dimensions before participating in the calculation.

[0066] The dual machine learning approach eliminates confounding bias through an orthogonalization step. In the first stage, a first gradient boosting tree model is fitted with the confounding variable X as input and the outcome variable Y as output. The first residual, the difference between the observed value Y and the model's predicted value, is calculated. This residual represents the fluctuation of the pure performance characteristics after excluding the effects of irradiance and temperature. In the second stage, a second gradient boosting tree model is fitted with the confounding variable X as input and the treatment variable D as output. The second residual, the difference between the observed value D and the model's predicted probability, is calculated. This residual represents the exogenous variation in the scheduling event that cannot be predicted by meteorological conditions. In the third stage, a linear regression without an intercept term is performed on the first residual and the second residual. The resulting regression coefficient represents the independent contribution after eliminating confounding bias; theoretically, this coefficient is equivalent to the average treatment effect.

[0067] To enhance the robustness of the estimation, a 5-fold cross-fit mechanism is adopted: all samples are divided into 5 folds, and two machine learning models are trained on each fold using the remaining 4 folds to predict the residual of the current fold. Finally, the residuals of all folds are summed for a third-stage regression. The hyperparameters of the machine learning model are: 200 trees, maximum depth of 6, learning rate of 0.05, and number of folds of 5. These parameters were determined through comparative experiments. Specifically, the search is performed on the candidate hyperparameter grid with the objective of minimizing the mean squared error of the validation set, selecting the combination with the lowest validation error.

[0068] The above process is executed for 12 scenarios combining 4 transmission performance characteristics and 3 scheduling event types, resulting in 12 independent contribution estimates.

[0069] For example, in a 50 MW photovoltaic power plant instance, power curtailment affects... The independent contribution is estimated to be approximately -0.11, meaning that the power curtailment order leads to an average reduction in transmission efficiency of approximately 0.11. This is consistent with engineering experience showing that the efficiency of an inverter deviates from its maximum power point during power curtailment, resulting in a decrease of 10% to 15%, thus validating the reasonableness of the estimate.

[0070] S5: Based on the independent contribution, remove the fluctuation component caused by the scheduling event from the performance feature vector sequence to obtain the purified device's own transmission performance degradation feature sequence.

[0071] In a specific embodiment of the present invention, step S5, the elimination operation specifically involves: determining whether each transmission performance feature value at each moment in the performance feature vector sequence falls within the time window of any scheduling event.

[0072] If it falls into the range, then subtract the independent contribution of the corresponding scheduling event to the transmission performance feature calculated in step S4 from the feature value.

[0073] If it does not fall into the category, the original eigenvalue remains unchanged.

[0074] After traversing all time points and all transmission performance characteristics, a purified degradation feature sequence is generated.

[0075] Specifically, step S5 uses the independent contribution values ​​from step S4 to systematically remove scheduling effects from the original performance feature vector sequence. For each transmission performance feature value at each time point, it is determined whether that time point falls within the time window of any scheduling event: if it falls within a single event window, the independent contribution value of the corresponding scheduling event to that feature value is subtracted from the feature value; if it falls within multiple overlapping event windows, the sum of the contributions of each overlapping event is subtracted, and the calculation process uses a linear superposition approximation method; if it does not fall within any window, the original value remains unchanged. After traversing all time points and all features, a purified degradation feature sequence is generated.

[0076] The verification method for the purification effect was as follows: a control sample with no confirmed equipment defects was selected, and its efficiency characteristics before and after power outages were compared. Experiments showed that in approximately 30 power outage events, the control sample initially showed better performance than the original sample. The average decrease during the power curtailment period was approximately 0.12. The difference between the power curtailment period and the non-power curtailment period after purification was... The mean difference was reduced to approximately 0.01, and about 92% of the scheduling effect was successfully removed.

[0077] For example, in a 50 MW photovoltaic power plant instance, before purification The sequence showed distinct dips in approximately 210 time frames across seven power outage events; these dips were effectively filled after purification. The sequence only showed a trend of decreasing by about 0.02 at the end of the analysis period, which is a true signal of equipment degradation.

[0078] It should be noted that the above-described elimination method based on constant independent contribution is suitable for scenarios where the intensity of scheduling events is relatively stable. In application scenarios where the intensity of scheduling events varies significantly, the independent contribution can be further modeled as a function of the intensity of scheduling events. That is, the elimination operation is performed after linearly or nonlinearly scaling the independent contribution according to the power curtailment ratio or reactive power regulation amplitude.

[0079] Specifically: For power rationing orders, define their intensity factor. Power rationing depth ratio: ,in This represents the available output at the moment the command is issued. To exert force for the objectives required by the command. The larger the value, the greater the depth of power rationing.

[0080] For reactive power regulation commands, define their intensity factor. The ratio of reactive power regulation amplitude to the inverter's rated reactive power capacity: ,in The target reactive power value is the command. The current reactive power output before the instruction is issued. This refers to the rated reactive power capacity of the inverter.

[0081] Average treatment effect estimated in step S4 based on intensity factor Scaling is performed to obtain the adjustment contribution of the scheduled event at the current intensity. The scaling method can be selected from the following: (a) Linear scaling: ,in This represents the actual intensity factor of the currently scheduled event. The reference intensity factor is taken as the mean of the intensity factors of all scheduling event samples used in the ATE estimation in step S4. This method assumes that the scheduling effect size is directly proportional to the scheduling intensity.

[0082] (b) Nonlinear scaling: Fitting the scheduling intensity factor using historical data Functional relationship between the corresponding effect size For example, by establishing local weighted regression or multinomial regression. The mapping curve between the effect size and the effect size is then This method can capture the nonlinear characteristics of the scheduling effect as a function of intensity.

[0083] Accordingly, in the elimination operation of step S5, for the moment that falls within the scheduling event time window, the adjusted contribution value after intensity scaling is subtracted from the feature value. rather than a fixed average treatment effect .

[0084] In scenarios where the intensity of scheduling events varies within a small range, such as power rationing fluctuations not exceeding ±10 percentage points, the error caused by constant contribution adjustment is negligible, and can be directly used. Simply removing the subtracted components will meet the engineering accuracy requirements; there is no need to introduce a strength scaling mechanism.

[0085] S6: Based on the purified degradation feature sequence, the transmission performance degradation index at the current moment is generated using a time-series prediction model, and the future degradation trend curve is extrapolated.

[0086] In a specific embodiment of the present invention, in step S6: the time-series prediction model is a long short-term memory network, whose input is a fixed-length sliding window of the purified degradation feature sequence, and whose output is the predicted degradation index value for multiple future time steps.

[0087] The transmission performance degradation index is a scalar index obtained by weighting and fusing the purified values ​​of active power transmission efficiency, reactive power response time, voltage deviation rate, and total harmonic distortion rate of current. Each weight is determined according to the order and magnitude of the abnormal responses of each transmission performance characteristic in historical fault samples.

[0088] Specifically, step S6 uses a long short-term memory network (LSTM) to generate a transmission performance degradation index based on the purified degradation feature sequence and extrapolates the future degradation trend curve.

[0089] The LSTM network input is a fixed-length sliding window containing purified degraded feature sequences. The window length was optimized through parameter comparison experiments, with candidate values ​​including 16, 32, 64, 128, and 256 steps, each corresponding to a duration of approximately 8 to 128 minutes. By comparing the root mean square error (RMSE) of predictions under different parameters, the 64-step model with the lowest error was ultimately selected, corresponding to an actual duration of approximately 32 minutes. The network structure is as follows: the input layer receives a tensor of 64 frames × 4 features; two LSTM hidden layers, each with 64 hidden units. During the research, various unit numbers within the range of 32 to 128 were compared, and 64 units achieved the optimal balance between model parameter count and data fitting ability; a fully connected layer maps the hidden states to a 16-step prediction output, with this output step count corresponding to a duration of 8 minutes; the output layer is linearly activated. The network is trained using an adaptive momentum estimation optimizer with RMSE as the loss function and a learning rate of 0.001. An early stopping mechanism is set to stop training if the verification loss does not decrease for 15 consecutive rounds.

[0090] Transmission performance degradation index It is a weighted fusion of four purification features, and after normalizing each feature to the [0,1] interval, it is calculated using the following formula:

[0091] Weights in the formula to The active power transmission efficiency was determined based on the order and distinguishability of abnormal responses in multiple historical fault cases for each transmission performance characteristic. The detectable anomaly appeared earliest in all four typical degradation modes, and was set to 0.4; Next, we take 0.3; then the voltage deviation rate, which is next, is 0.2; and finally, the reactive power response time, which is the latest and only sensitive to inverter-related degradation, is 0.1. The closer the degradation index is to 1, the more severe the degradation.

[0092] The training labels are a sequence of degradation indices calculated using the real cleansing features of the next 16 steps with equal weights. During inference, the first value of the output sequence is taken as the current degradation index, and the entire sequence is connected to form an extrapolation trend curve. A Monte Carlo uncertainty estimation method based on Dropout is used to generate prediction confidence intervals. This method sets a dropout rate of 0.2 and repeats the forward propagation operation 30 times.

[0093] For example, in a 50 MW photovoltaic power plant instance, the first 5000 frames were used for training, followed by 760 frames for testing. In frame 280 of the test set, the degradation index climbed from approximately 0.15 at the baseline to 0.34, and the extrapolated trend indicated it would continue to climb to approximately 0.42 within the next 8 minutes. A subsequent inspection 48 hours later revealed a bulging DC-side filter capacitor on a 500 kW inverter, verifying the accuracy of the prediction. During the previous 10 hours of power curtailment, the original... Despite significant fluctuations, the purification and degradation index has remained stable.

[0094] It should be noted that for scenarios where historical data is insufficient in the early stages of deployment, the following cold start strategy can be adopted: use the transfer learning pre-trained model of the same type of photovoltaic power station in the same region as the initial model, and fine-tune it gradually as the data of this power station accumulates; or use the degradation index calculation based on the physical degradation model as the initial supervision signal of the LSTM model, and switch to LSTM prediction mode after the historical data accumulates to a preset scale, such as 1440 frames, about 12 hours later.

[0095] S7: Compare the transmission performance degradation index with preset multi-level thresholds. When the warning threshold is exceeded, generate an alarm and output a performance evaluation report containing a decoupling explanation of the scheduling event.

[0096] Specifically, the multi-level thresholds include a first warning threshold, a second alarm threshold, and a third fault threshold. Each threshold is determined based on the statistical distribution of the degradation index under normal operating conditions in historical data. For example, the first warning threshold is the mean of the historical baseline value of the degradation index plus two standard deviations, corresponding to an anomaly detection probability of approximately 95%; the second alarm threshold is the mean plus three standard deviations; and the third fault threshold is the mean plus four standard deviations.

[0097] In a specific embodiment of the present invention, the method further includes: step S8: when step S7 determines that the warning threshold has not been exceeded, the transmission performance degradation index generated in step S6 is used as a feedback signal to update the estimated parameters of the causal inference algorithm in step S4 and the network weights of the time series prediction model in step S6, forming a closed-loop adaptive optimization mechanism from the evaluation result to causal decoupling and degradation prediction.

[0098] Specifically, when the degradation index does not exceed any warning threshold, data from the normal stable period is used for closed-loop optimization.

[0099] The first feedback to step S4: During a continuous stable period without scheduling events, examine whether the purified features still exhibit structured residual fluctuations related to irradiance or temperature. If the absolute value of the partial correlation coefficient exceeds approximately 0.2, it indicates that the existing set of confounding variables is incomplete, and variables such as the number of operating inverters and the degree of dust accumulation in the photovoltaic array need to be added, and the independent contribution needs to be re-estimated.

[0100] The second feedback loop leads to step S6: For every 1440 steps (approximately 12 hours) of newly added normal-state degradation index data, the LSTM model undergoes 5 rounds of fine-tuning training, with the learning rate reduced to one-tenth of the initial value. This allows the model to adapt to the latest data distribution while retaining knowledge of historical degradation patterns. The update cycle is determined based on the average time span from normal to detectable degradation, which is approximately 120 hours; 12 hours is much shorter than this scale.

[0101] For example, during the first 40 hours of the stable period in a 50 MW photovoltaic power plant instance, after purification The partial correlation coefficient with irradiance is 0.14. The partial correlation coefficient with component temperature was 0.18, all below the 0.2 threshold, indicating that the existing set of confusing variables was sufficient. Three incremental LSTM adjustments were performed during the process, and the prediction error remained stable throughout subsequent degradation detection.

[0102] Reference Figure 2 As shown, the second aspect of the present invention discloses a system for performing the power dispatch-based photoelectric transmission performance evaluation method described in the present invention, comprising: a data acquisition module for synchronously acquiring and time-aligning real-time operating data of photovoltaic power plants, historical dispatch command data issued by the power grid dispatch center, and environmental meteorological data of the power plant, and generating a multi-source fusion dataset.

[0103] The feature extraction module is used to extract transmission performance features, including active power transmission efficiency, reactive power response time, voltage deviation rate, and total harmonic distortion rate of current, based on a multi-source fusion dataset, and construct a time-stamped performance feature vector sequence.

[0104] The scheduling event parsing module is used to parse historical scheduling instruction data, identify and mark scheduling events of the types of power restriction instructions, peak shaving instructions and reactive power regulation instructions, and generate a scheduling event tag sequence.

[0105] The causal decoupling calculation module is used to temporally correlate the performance feature vector sequence with the scheduling event label sequence. For each type of scheduling event, the causal inference algorithm is used to calculate the independent contribution of each scheduling event to each transmission performance feature. Based on the independent contribution, the fluctuation component caused by the scheduling event is removed from the performance feature vector sequence to obtain the purified equipment transmission performance degradation feature sequence, wherein at least the irradiance and temperature in the field environmental meteorological data are used as confounding variables.

[0106] The degradation prediction module is used to generate the transmission performance degradation index at the current moment and extrapolate the future degradation trend curve based on the purified degradation feature sequence using a time-series prediction model.

[0107] The evaluation and alarm module compares the transmission performance degradation index with preset multi-level thresholds. When the warning threshold is exceeded, an alarm is generated, and a performance evaluation report containing a decoupled explanation of the scheduling event is output.

[0108] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0109] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0110] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for evaluating optical-electrical transmission performance based on power scheduling, characterized in that, include: S1: Synchronously collect and time-align real-time operation data of photovoltaic power plants, historical dispatch instructions issued by the power grid dispatch center, and environmental meteorological data of the power plants to generate multi-source fusion datasets; S2: Based on the multi-source fusion dataset, extract transmission performance features including active power transmission efficiency, reactive power response time, voltage deviation rate and total harmonic distortion rate of current, and construct a time-stamped performance feature vector sequence. S3: Parse historical dispatch instruction data, identify and label dispatch events of power restriction instruction, peak shaving instruction and reactive power regulation instruction types, and generate a dispatch event tag sequence; S4: Correlate the performance feature vector sequence with the scheduling event label sequence in time. For each type of scheduling event, use a causal inference algorithm to calculate the independent contribution of each scheduling event to each transmission performance feature, with at least irradiance and temperature in the station's environmental meteorological data as confounding variables. S5: Based on the independent contribution, remove the fluctuation component caused by the scheduling event from the performance feature vector sequence to obtain the purified equipment's own transmission performance degradation feature sequence; S6: Based on the purified degradation feature sequence, the transmission performance degradation index at the current moment is generated using a time-series prediction model, and the future degradation trend curve is extrapolated. S7: Compare the transmission performance degradation index with preset multi-level thresholds. When the warning threshold is exceeded, generate an alarm and output a performance evaluation report containing a decoupling explanation of the scheduling event.

2. The method of claim 1, wherein, In step S1: The real-time operating data includes the DC side voltage and current of the photovoltaic array, the AC side output power of the inverter, the voltage and current of each node of the collector line, and the active and reactive power at the grid connection point. The historical dispatch instruction data includes power rationing instructions issued by AGC and their execution time windows, reactive power voltage regulation instructions issued by AVC and their execution time windows, as well as planned output curves; The environmental meteorological data of the site includes total horizontal irradiance, component backplane temperature, ambient temperature, and wind speed.

3. The method of claim 1, wherein the method further comprises: In step S2: The active power transmission efficiency is the ratio of the active power output at the grid connection point to the theoretical maximum DC power of the photovoltaic array calculated based on real-time irradiance and component temperature. The reactive power response time is the time elapsed from the moment the reactive power adjustment command is issued until the inverter's reactive power output first reaches a preset percentage of the command target value. The voltage deviation rate is the absolute percentage of the effective value of the grid connection voltage deviating from the rated voltage; The total harmonic distortion rate of the current is the ratio of the square root of the sum of the squares of the effective values ​​of each harmonic component in the grid-connected current to the effective value of the fundamental component.

4. The method of claim 1, wherein, The parsing of historical scheduling instruction data in step S3 specifically includes: Parse the instruction type field in the dispatch instruction message to distinguish between power rationing instructions, peak shaving instructions, and reactive power regulation instructions; Extract the start and end times of the execution time window corresponding to each instruction; Multiple scheduling events with overlapping time windows are merged, and different types of scheduling events with overlapping time windows are retained and marked with overlap to generate a sequence of scheduling event labels.

5. The method of claim 1, wherein, In step S4, the causal inference algorithm is a causal effect estimation algorithm based on the counterfactual reasoning framework. It takes each transmission performance feature in the performance feature vector sequence as the result variable, the scheduling event as the processing variable, and the irradiance and temperature in the station environmental meteorological data as at least the confounding variables. It estimates the average processing effect of each scheduling event on each transmission performance feature and takes the average processing effect as the independent contribution.

6. The method of claim 5, wherein the method further comprises: The causal effect estimation algorithm based on the counterfactual reasoning framework is a dual machine learning method. This method estimates the independent contribution by fitting a first machine learning model with confounding variables as input and transmission performance features as output, and calculating a first residual, which is the difference between the observed value of the transmission performance feature and the predicted value of the first machine learning model. The second machine learning model is fitted with the confounding variable as input and the binary indicator variable of the scheduling event as output, and the second residual is calculated as the difference between the observed value of the scheduling event indicator variable and the predicted value of the second machine learning model. Regressing the first residual against the second residual yields regression coefficients that represent the independent contributions after eliminating confounding bias.

7. The method of claim 5, wherein the method further comprises: The confounding variables also include the number of inverters in operation and the degree of dust accumulation in the photovoltaic array. The degree of dust accumulation is obtained by comparing the difference in DC-side output power of the array under different cleaning cycles under the same irradiance conditions. 8.The power scheduling based optical-electrical transmission performance evaluation method of claim 1, wherein, In step S5, the rejection operation specifically includes: For each transmission performance feature value at each moment in the performance feature vector sequence, determine whether that moment falls within the time window of any scheduling event. If it falls into the range, subtract the independent contribution of the corresponding scheduling event to the transmission performance feature calculated in step S4 from the feature value. If it does not fall into the category, the original eigenvalue remains unchanged; After traversing all time points and all transmission performance characteristics, a purified degradation feature sequence is generated.

9. The method for evaluating the performance of photoelectric transmission based on power dispatch according to claim 1, characterized in that, In step S6: the time-series prediction model is a long short-term memory network, whose input is a fixed-length sliding window of the purified degradation feature sequence, and whose output is the predicted degradation index value for multiple future time steps. The transmission performance degradation index is a scalar index obtained by weighting and fusing the purified values ​​of active power transmission efficiency, reactive power response time, voltage deviation rate, and total harmonic distortion rate of current. Each weight is determined according to the order and magnitude of the abnormal responses of each transmission performance characteristic in historical fault samples.

10. A system for performing the power dispatch-based photoelectric transmission performance evaluation method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to synchronously collect and time-align real-time operation data of photovoltaic power plants, historical dispatch command data issued by the power grid dispatch center, and environmental meteorological data of the power plant, and generate multi-source fusion datasets. The feature extraction module is used to extract transmission performance features, including active power transmission efficiency, reactive power response time, voltage deviation rate and total harmonic distortion rate of current, based on a multi-source fusion dataset, and construct a time-stamped performance feature vector sequence. The scheduling event parsing module is used to parse historical scheduling instruction data, identify and mark scheduling events of the types of power restriction instructions, peak shaving instructions and reactive power adjustment instructions, and generate a scheduling event tag sequence; The causal decoupling calculation module is used to temporally correlate the performance feature vector sequence with the scheduling event label sequence. For each type of scheduling event, the causal inference algorithm is used to calculate the independent contribution of each scheduling event to each transmission performance feature. Based on the independent contribution, the fluctuation component caused by the scheduling event is removed from the performance feature vector sequence to obtain the purified equipment transmission performance degradation feature sequence, wherein at least the irradiance and temperature in the field environmental meteorological data are used as confounding variables. The degradation prediction module is used to generate the transmission performance degradation index at the current moment and extrapolate the future degradation trend curve based on the purified degradation feature sequence using a time series prediction model. The evaluation and alarm module compares the transmission performance degradation index with preset multi-level thresholds. When the warning threshold is exceeded, an alarm is generated, and a performance evaluation report containing a decoupled explanation of the scheduling event is output.

Images