A machine learning based distributed photovoltaic short-term power generation prediction system

Abstract

The application discloses a kind of distributed photovoltaic short-term power generation power prediction systems based on machine learning, including main terminal, light-heat coaxial acquisition module, acoustic vibration synchronous detection module and photovoltaic short-term power generation power prediction module;The light-heat coaxial acquisition module, acoustic vibration synchronous detection module are connected with main terminal communication respectively, and the photovoltaic short-term power generation power prediction module is connected with main terminal communication;The light-heat coaxial acquisition module collects real-time light intensity data and photovoltaic module temperature data, and is transmitted to main terminal;Photovoltaic short-term power generation power prediction module uses hierarchical fusion model, combines physical basic layer and data driven layer, dynamically optimizes weight, and introduces power attenuation coefficient to correct, significantly improves prediction accuracy and adaptability.System also integrates multi-source data collaborative access unit, ensures data reliability through data quality closed-loop management, effectively deals with data heterogeneous problem under distributed scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of photovoltaic prediction, more particularly to a distributed photovoltaic short-term power generation prediction system based on machine learning. BACKGROUND

[0002] With the continuous expansion of distributed photovoltaic power generation, accurate prediction of its short-term power generation is crucial for grid dispatching, energy management and stable operation.

[0003] However, existing prediction systems rely on a single data source or simple models, which have obvious limitations: on the one hand, traditional physical models do not fully consider the power attenuation caused by the actual state of the equipment (such as aging, failure), while data-driven models are sensitive to data quality and lack physical constraints; on the other hand, distributed photovoltaic stations are scattered, data is heterogeneous, and weather conditions and equipment operating conditions are variable, making it difficult to guarantee prediction accuracy. In addition, existing systems often lack multi-dimensional data coordination and fusion mechanisms, such as simultaneous detection of light intensity, temperature and acoustic vibration signals, as well as linkage between prediction results and operation and maintenance management, limiting practicality and reliability. Therefore, there is an urgent need for an intelligent prediction system that can integrate multi-source data, fuse physical and machine learning methods, and monitor equipment status in real time, to improve the accuracy and robustness of distributed photovoltaic short-term power prediction. SUMMARY

[0004] To solve the above problems, the present application provides a distributed photovoltaic short-term power generation prediction system based on machine learning.

[0005] The present application provides a distributed photovoltaic short-term power generation prediction system based on machine learning, which comprises a main terminal, a light-heat coaxial collection module, an acoustic vibration synchronous detection module and a photovoltaic short-term power generation prediction module; the light-heat coaxial collection module and the acoustic vibration synchronous detection module are respectively in communication connection with the main terminal, and the photovoltaic short-term power generation prediction module is in communication connection with the main terminal.

[0006] The light-heat coaxial collection module collects real-time light intensity data and photovoltaic module temperature data and transmits them to the main terminal; the acoustic vibration synchronous detection module collects acoustic vibration signals of photovoltaic equipment, identifies acoustic vibration abnormalities and converts them into power attenuation coefficients γ and transmits them to the main terminal.

[0007] The main terminal integrates the received data and transmits it to the photovoltaic short-term power generation prediction module; the photovoltaic short-term power generation prediction module calculates the photovoltaic short-term power generation based on multi-dimensional sensing data and power attenuation coefficients γ through a hierarchical fusion prediction model, outputs the prediction results and related evaluation indicators and returns them to the main terminal.

[0008] Preferably, the photovoltaic short-term power generation prediction module receives multi-dimensional sensor data transmitted from the main terminal, performs prediction calculations of photovoltaic short-term power generation based on this data, and sends the prediction results and related evaluation indicators back to the main terminal.

[0009] The evaluation indicators include Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE), and their calculation formulas are as follows:

[0010]

[0011] in Let the predicted power be at time k. Let n be the actual power at time k, and n be the number of sampling points in the prediction period.

[0012] ;

[0013] .

[0014] Preferably, when the photovoltaic short-term power generation prediction module performs prediction calculations, it first preprocesses the input multi-dimensional sensor data, including outlier removal based on the 3σ criterion and missing value filling based on linear interpolation.

[0015] Then, a hierarchical fusion prediction model is used, which includes:

[0016] Physical foundation layer: The theoretical power value is calculated based on the formula PPHYS=PSTC×(G / GSTC)×[1+αT×(Tc-TSTC)], where PSTC is the rated power under standard test conditions, G is the real-time light intensity, GSTC is the standard light intensity, αT is the power temperature coefficient, Tc is the component temperature, and TSTC is the standard temperature.

[0017] Data-driven layer: A bidirectional LSTM network with an improved attention mechanism is used. The input features include time series of light intensity and temperature and historical power data. The output correction coefficient δ satisfies PML=PPHYS×δ.

[0018] Dynamic fusion layer: The final prediction is achieved through the formula PFINAL=w1×PPHYS+w2×PML, where w1+w2=1;

[0019] It should be noted that (w1) and (w2) are dynamic weights based on historical prediction errors, which are obtained by iterative optimization using the least squares method, i.e., minimizing \(sum{k=1}^{m}(P{text{FINAL},k} and P{text{real},k})^2, where m is the number of historical training samples. The smaller the historical error, the larger the weight of the layer.

[0020] The prediction module simultaneously outputs a 95% confidence interval of the prediction result, satisfying CI=[PFINAL and 1.96×σ, PFINAL+1.96×σ], wherein σ is a standard deviation of a prediction error.

[0021] Preferably, the photo-thermal coaxial acquisition module comprises a visible light component and an infrared component arranged coaxially, both of which realize exposure timing synchronization through a programmable logic unit and satisfy a corrected exposure time formula t3=k1×t0×(GSTC / Greal);

[0022] wherein k1 is a type coefficient adapted to different photovoltaic related devices, t0 is a reference exposure time, GSTC is a standard light intensity, and G is a real-time light intensity.

[0023] The infrared component outputs device temperature data, and the visible light component outputs real-time light intensity data, both of which realize spatial alignment through an improved ORB feature point matching technology, and a curved surface projection correction step is added in view of the curved surface installation characteristics of the photovoltaic component to ensure that the temperature data and the light intensity data correspond to the same component region.

[0024] The photo-thermal coaxial acquisition module transmits the processed light intensity data and temperature data to the main terminal as core input parameters of photovoltaic short-term power generation prediction.

[0025] The acoustic vibration synchronous detection module comprises a multi-array ultrasonic array and a three-axis vibration sensor, the multi-array ultrasonic array realizes directional focusing of ultrasonic signals through beam focusing, and the array spacing of the focal points is designed in view of the partial discharge characteristics of the photovoltaic inverter to improve the capture accuracy of the partial discharge signals.

[0026] The vibration sensor collects device vibration time domain signals, the time domain signals are converted into frequency domain spectra through Fourier transform, and after extracting specific frequency band features related to photovoltaic device faults (such as 10 kHz and 50 kHz frequency bands corresponding to inverter partial discharge), difference analysis is performed with a preset normal frequency spectrum model; when the difference satisfies a set threshold, it is determined as an acoustic vibration abnormality.

[0027] The acoustic vibration synchronous detection module transmits acoustic vibration abnormality information to the main terminal and converts it into a power attenuation coefficient γ according to the acoustic vibration abnormality level, which is used to correct the photovoltaic short-term power generation prediction value.

[0028] It should be noted that the vibration sensor collects device vibration time domain signals, the time domain signals are converted into frequency domain spectra through Fourier transform, and after extracting specific frequency band features related to photovoltaic device faults (such as 10 kHz and 50 kHz frequency bands corresponding to inverter partial discharge), difference analysis is performed with a preset normal frequency spectrum model; the "set threshold" is obtained based on the acoustic vibration spectrum statistics of normal devices, i.e. the mean value (mu) of the spectrum amplitude of the normal device in this frequency band ±3σ (σ is the standard deviation of the spectrum amplitude of the normal device), when the difference exceeds (mu+3σ), it is determined as an acoustic vibration abnormality.

[0029] Preferably, it further comprises a multi-source data collaborative access unit, which is in communication connection with the main terminal;

[0030] The multi-source data collaborative access unit integrates a CMS system, a SCADA mirror system, a marketing system, a business middle platform and a weather prediction system, and realizes the timed or real-time extraction of account data, NWP predicted weather data, real-time weather data, real-time power data and running state data through a data access engine; the multi-source data collaborative access unit performs preliminary quality screening on the collected data, removes data obviously beyond the physical characteristics of the photovoltaic module, and then transmits the preprocessed data to the main terminal, thereby providing a data basis for the model training, optimization and evaluation of the photovoltaic short-term power generation prediction module.

[0031] Preferably, the preprocessing further comprises a hierarchical data quality closed-loop management process, specifically:

[0032] First step, data auditing: a three-level auditing mechanism of format verification, logic verification and physical verification is constructed, the format verification ensures that the data conforms to the preset field specification, the logic verification verifies the consistency of the associated data, and the physical verification removes the out-of-limit data based on the rated parameters of the photovoltaic module;

[0033] Second step, abnormal classification and tracing: the abnormal data is classified into three types of missing type, mutation type and deviation type, the abnormal reasons are located through data tracing technology, and a data abnormality tracing report is generated;

[0034] Third step, differential repair: customized repair strategies are adopted for different types of abnormalities, the missing data is interpolated with historical similar weather data, the mutation data is smoothed with a sliding window and corrected with adjacent period trend, and the deviation data is corrected based on the Pearson correlation coefficient of weather and power, so as to ensure that the repaired data matches the actual weather conditions;

[0035] Fourth step, quality feedback: the data qualification rate is calculated and associated with the prediction accuracy, a closed-loop feedback of data quality and prediction effect is formed, and the auditing and repair strategies are dynamically optimized.

[0036] Preferably, 7. The photovoltaic short-term power generation prediction module is designed for distributed photovoltaic scenarios, and a dynamic benchmark station cluster and a multi-factor expansion fusion prediction scheme are designed, and the specific process is as follows:

[0037] First step, dynamic construction of benchmark station cluster: based on the screening conditions of geographical location, climate similarity and data qualification rate, the distributed photovoltaic power stations in the region are screened in real time to form a benchmark station cluster, and the cluster members are updated regularly based on the latest weather data and equipment status, and the faulty and repaired power stations are removed.

[0038] Second step, cluster-level model training: taking the total actual power of the reference station cluster as the prediction target, inputting regional NWP predicted meteorological data, cluster historical power data, and real-time light intensity and temperature data of the light-heat coaxial collection module, using an improved attention mechanism bidirectional LSTM and XGBoost integrated model to construct the meteorological and power mapping relationship, and outputting the cluster-level short-term predicted power;

[0039] Third step, multi-factor expansion calculation: for the distributed photovoltaic power station with incomplete data in the region, meteorological similarity factor, equipment aging factor, and installed capacity factor are introduced to construct an expansion model, and the predicted power is calculated through the expansion model;

[0040] Fourth step, result correction: combining the power attenuation coefficient γ of the sound vibration synchronous detection module, the predicted power obtained by the expansion calculation is corrected to obtain the final predicted power.

[0041] It should be noted that in the third step, multi-factor expansion calculation: for the distributed photovoltaic power station with incomplete data in the region, meteorological similarity factor S, equipment aging factor A, and installed capacity factor C are introduced to construct an expansion model, and the predicted power formula is \(P_{text{target}}=P_{text{cluster}}\times\frac{C_{text{target}}}{C_{text{cluster}}}\times S\times A); wherein \(C_{text{target}}) is the installed capacity of the target power station, \(C_{text{cluster}}) is the total installed capacity of the reference station cluster; S is the meteorological similarity of the target power station and the cluster (calculated by the Pearson correlation coefficient of light intensity and temperature, taking values 0.8 and 1.0); A is the equipment aging factor of the target power station (based on service life fitting, when serving t years \(A=1 and 0.01t), t≤10).

[0042] Preferably, the evaluation index further includes correlation coefficient r, accuracy Cr, and qualified rate Qr; the calculation formula of the correlation coefficient r is:

[0043] , wherein P1 is the average value of the actual power in the statistical period, and P2 is the average value of the predicted power in the statistical period;

[0044] The calculation formula of the accuracy Cr is: , wherein is the on-line capacity of the photovoltaic equipment at the kth moment; the calculation formula of the qualified rate Qr is: , wherein is the determination result.

[0045] It should be noted that;

[0046] When ((|Pk^{text{pre}}andPk^{text{actual}}| / Pk^{text{actual}}))≤5%, \(Bk=1) (qualified as qualified), otherwise \(Bk=0) (qualified as unqualified).

[0047] Preferably, the main terminal is also used for predictive output correction, generating correction coefficients based on maintenance plans, power curtailment plans, equipment status, and installed capacity data, specifically:

[0048] The system obtains the maintenance plan of the photovoltaic power station and generates maintenance correction coefficients; it also obtains the power curtailment plan of the power grid and generates curtailment correction coefficients; and combines the power attenuation coefficient γ output by the acoustic vibration synchronous detection module to correct the prediction results using the overall correction formula, thus obtaining the corrected final predicted power.

[0049] It should be noted that the overall correction formula is:

[0050] Obtain the maintenance plan of the photovoltaic power station and generate the maintenance correction coefficient \(alpha) (\(alpha=1 and \beta) when the installed capacity of the maintenance equipment is \(beta)); obtain the power curtailment plan of the power grid and generate the curtailment correction coefficient \(delta) (\(delta=1 and \gamma{text{limit}}) when the curtailment ratio is \(gamma{text{limit}})); combine the power attenuation coefficient \(gamma) output by the acoustic vibration synchronous detection module;

[0051] By using the general correction formula:

[0052] The prediction result is corrected by \(P{text{final}}=P{text{predicted}}\times\alpha\times\delta\times(1 and\gamma)) to obtain the corrected final predicted power; where \(P{text{predicted}}) is the initial predicted power output by the photovoltaic short-term power generation prediction module.

[0053] Beneficial effects:

[0054] By using a photothermal coaxial acquisition module and an acoustic vibration synchronous detection module, high-precision synchronous acquisition and anomaly identification of real-time light intensity, temperature and equipment acoustic vibration signals are achieved, providing multi-dimensional and high-quality input data for power prediction.

[0055] The photovoltaic short-term power generation prediction module adopts a hierarchical fusion model, combining a physical foundation layer and a data-driven layer, dynamically optimizing weights, and introducing a power attenuation coefficient for correction, significantly improving prediction accuracy and adaptability. The system also integrates a multi-source data collaborative access unit, ensuring data reliability through closed-loop data quality management, effectively addressing data heterogeneity issues in distributed scenarios.

[0056] Furthermore, the dynamic reference station cluster and multi-factor expansion scheme enhance the generalization ability of predictions, while the prediction-operation and maintenance linkage visualization module supports power curve comparison, penetrating analysis, and intelligent report generation, realizing the coordination of prediction results with equipment status and maintenance plans, thus contributing to grid dispatch optimization and improved operation and maintenance efficiency. Overall, the system not only reduces prediction errors (e.g., through quantification using evaluation indicators RMSE, MAE, and MAPE), but also enhances the system's practicality and scalability, providing a comprehensive solution for the intelligent management of distributed photovoltaic systems. Attached Figure Description

[0057] Figure 1 This is a flowchart of the generation method of the present invention. Detailed Implementation

[0058] like Figure 1 As shown: A distributed photovoltaic short-term power generation prediction system based on machine learning includes a main terminal, a solar thermal coaxial acquisition module, an acoustic vibration synchronous detection module, and a photovoltaic short-term power generation prediction module; the solar thermal coaxial acquisition module and the acoustic vibration synchronous detection module are respectively communicatively connected to the main terminal, and the photovoltaic short-term power generation prediction module is also communicatively connected to the main terminal.

[0059] The coaxial solar thermal acquisition module collects real-time light intensity data and photovoltaic module temperature data, and transmits them to the main terminal; the synchronous acoustic and vibration detection module collects acoustic and vibration signals of the photovoltaic equipment, identifies acoustic and vibration anomalies, and converts them into a power attenuation coefficient γ, which is then transmitted to the main terminal.

[0060] After integrating the received data, the main terminal transmits it to the photovoltaic short-term power generation prediction module. The photovoltaic short-term power generation prediction module calculates the photovoltaic short-term power generation based on multi-dimensional sensor data and power attenuation coefficient γ through a hierarchical fusion prediction model, outputs the prediction results and related evaluation indicators, and sends them back to the main terminal.

[0061] As an optional embodiment, the photovoltaic short-term power generation prediction module receives multi-dimensional sensing data transmitted from the main terminal, performs prediction calculations of photovoltaic short-term power generation based on this data, and sends the prediction results and related evaluation indicators back to the main terminal.

[0062] The evaluation indicators include Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE), and their calculation formulas are as follows:

[0063]

[0064] in Let the predicted power be at time k. Let n be the actual power at time k, and n be the number of sampling points in the prediction period.

[0065] ;

[0066] .

[0067] As an optional embodiment, when the photovoltaic short-term power generation prediction module performs prediction calculations, it first preprocesses the input multi-dimensional sensing data, including outlier removal based on the 3σ criterion and missing value filling based on linear interpolation.

[0068] Then, a hierarchical fusion prediction model is used, which includes:

[0069] Physical foundation layer: The theoretical power value is calculated based on the formula PPHYS=PSTC×(G / GSTC)×[1+αT×(Tc and TSTC)], where PSTC is the rated power under standard test conditions, G is the real-time light intensity, GSTC is the standard light intensity, αT is the power temperature coefficient, Tc is the component temperature, and TSTC is the standard temperature.

[0070] Data-driven layer: A bidirectional LSTM network with an improved attention mechanism is used. The input features include time series of light intensity and temperature and historical power data. The output correction coefficient δ satisfies PML=PPHYS×δ.

[0071] Dynamic fusion layer: The final prediction is achieved through the formula PFINAL=w1×PPHYS+w2×PML, where w1+w2=1;

[0072] It should be noted that \(w1) and \(w2) are dynamic weights based on historical prediction errors, which are obtained by iterative optimization using the least squares method, i.e., minimizing \(sum{k=1}^{m}(P{text{FINAL},k} and P{text{real},k})^2), where m is the number of historical training samples. The smaller the historical error, the larger the weight of the layer.

[0073] The prediction module also outputs the 95% confidence interval of the prediction result, which satisfies CI=[PFINAL and 1.96×σ,PFINAL+1.96×σ], where σ is the standard deviation of the prediction error.

[0074] As an optional embodiment, the photothermal coaxial acquisition module includes a visible light component and an infrared component arranged coaxially. The two are synchronized in exposure timing through a programmable logic unit and satisfy the corrected exposure time formula t3=k1×t0×(GSTC / Greal).

[0075] Where k1 is the type coefficient for adapting to different photovoltaic-related equipment, t0 is the reference exposure time, GSTC is the standard light intensity, and G is the real-time light intensity;

[0076] The infrared component outputs temperature data, and the visible light component outputs real-time light intensity data. The two types of data are spatially aligned using an improved ORB feature point matching technology. A curved surface projection correction step is added to address the curved surface installation characteristics of the photovoltaic module, ensuring that the temperature data and light intensity data correspond to the same module area.

[0077] The solar thermal coaxial acquisition module transmits the processed light intensity data and temperature data to the main terminal, which serves as the core input parameters for short-term photovoltaic power generation prediction.

[0078] The acoustic and vibration synchronous detection module includes a multi-element ultrasonic array and a triaxial vibration sensor. The multi-element ultrasonic array achieves directional focusing of ultrasonic signals through beam focusing. The focal point array spacing is designed for the partial discharge characteristics of photovoltaic inverters to improve the accuracy of partial discharge signal capture.

[0079] The vibration sensor collects the vibration time-domain signal of the equipment. The time-domain signal is converted into a frequency domain spectrum by Fourier transform. After extracting specific frequency band features related to photovoltaic equipment faults, the difference is analyzed with the preset normal spectrum model. When the difference meets the set threshold, it is determined to be an acoustic vibration abnormality.

[0080] The acoustic and vibration synchronous detection module transmits acoustic and vibration anomaly information to the main terminal and converts it into a power attenuation coefficient γ based on the acoustic and vibration anomaly level, which is used to correct the short-term photovoltaic power generation prediction value.

[0081] It should be noted that the vibration sensor collects the vibration time-domain signal of the equipment. The time-domain signal is converted into a frequency domain spectrum through Fourier transform. After extracting specific frequency band features related to photovoltaic equipment faults (such as the 10kHz and 50kHz frequency bands corresponding to partial discharge of the inverter), a difference analysis is performed with the preset normal spectrum model. The "set threshold" is based on the statistical analysis of the acoustic and vibration spectrum of normal equipment, that is, the mean spectral amplitude of normal equipment in this frequency band is \(mu)±3σ (σ is the standard deviation of the spectral amplitude of normal equipment). When the difference exceeds \(mu+3σ), it is judged as acoustic and vibration abnormality.

[0082] As an optional embodiment, it also includes a multi-source data collaborative access unit, which is communicatively connected to the main terminal;

[0083] The multi-source data collaborative access unit integrates the CMS system, SCADA mirror system, marketing procurement system, business middleware, and weather forecasting system. Through the data access engine, it realizes the timed or real-time extraction of ledger data, NWP forecast weather data, real-time weather data, real-time power data, and operating status data. The multi-source data collaborative access unit performs preliminary quality screening on the collected data, removes data that obviously exceeds the physical characteristics of photovoltaic modules, and then transmits the pre-processed data to the main terminal, providing a data foundation for the model training, optimization, and evaluation of the photovoltaic short-term power generation prediction module.

[0084] As an optional embodiment, preprocessing also includes a hierarchical data quality closed-loop management process, specifically:

[0085] The first step is data review: a three-level review mechanism of format verification, logical verification and physical verification is established. Format verification ensures that the data conforms to the preset field specifications, logical verification verifies the consistency of related data, and physical verification removes data that exceeds the limit based on the rated parameters of photovoltaic modules.

[0086] The second step is anomaly classification and source tracing: the abnormal data is classified into three categories: missing, mutation, and deviation. The cause of the anomaly is located through data source tracing technology, and a data anomaly source tracing report is generated.

[0087] The third step is differentiated repair: customized repair strategies are adopted for different types of anomalies. Missing data are interpolated using historical similar meteorological day data, abrupt data are smoothed using a sliding window and corrected for trends in adjacent time periods, and deviation data are corrected inversely based on the meteorological and power Pearson correlation coefficients to ensure that the repaired data matches the actual meteorological conditions.

[0088] The fourth step is quality feedback: calculate the data pass rate and correlate it with the prediction accuracy to form a closed-loop feedback between data quality and prediction effect, and dynamically optimize the review and remediation strategies.

[0089] As an optional embodiment, 7. The photovoltaic short-term power generation prediction module, for distributed photovoltaic scenarios, designs a dynamic reference station cluster and a multi-factor expansion and fusion prediction scheme, the specific process of which is as follows:

[0090] The first step is to dynamically construct the benchmark station cluster: using geographical location, climate similarity, and data qualification rate as screening criteria, distributed photovoltaic power stations in the region are selected in real time to form a benchmark station cluster. The cluster members are updated regularly based on the latest meteorological data and equipment status, and faulty power stations are removed and those under maintenance are repaired.

[0091] The second step is cluster-level model training: taking the total actual power of the base station cluster as the prediction target, the input is regional NWP prediction meteorological data, cluster historical power data and real-time light intensity and temperature data of the photothermal coaxial acquisition module. The improved attention mechanism bidirectional LSTM and XGBoost integrated model are used to construct the meteorological and power mapping relationship and output the cluster-level short-term predicted power.

[0092] The third step is multi-factor capacity expansion calculation: For distributed photovoltaic power stations with incomplete data in the region, meteorological similarity factor, equipment aging factor, and installed capacity factor are introduced to build an expansion model, and the predicted power is calculated through the expansion model.

[0093] The fourth step is result correction: the predicted power obtained from the capacity expansion calculation is corrected by combining the power attenuation coefficient γ of the acoustic vibration synchronization detection module to obtain the final predicted power.

[0094] It should be noted that in the third step, multi-factor capacity expansion calculation: for distributed photovoltaic power stations with incomplete data in the region, a capacity expansion model is constructed by introducing a meteorological similarity factor S, an equipment aging factor A, and an installed capacity factor C. The predicted power formula is \(P_{text{target}}=P_{text{cluster}}\times\frac{C_{text{target}}}{C_{text{cluster}}}\times S\times A); where \(C_{text{target}}) is the installed capacity of the target power station, and \(C_{text{cluster}}) is the total installed capacity of the base station cluster; S is the meteorological similarity between the target power station and the cluster (calculated by the Pearson correlation coefficient of light intensity and temperature, with values ​​of 0.8 and 1.0); A is the equipment aging factor of the target power station (fitted based on the service life, \(A=1 and 0.01t) when the service life is t years, t≤10).

[0095] As an optional embodiment, the evaluation indicators also include the correlation coefficient r, accuracy Cr, and pass rate Qr; the formula for calculating the correlation coefficient r is:

[0096] , where P1 is the average actual power during the statistical period, and P2 is the average predicted power during the statistical period;

[0097] The formula for calculating accuracy Cr is: ,in Let be the photovoltaic equipment's operating capacity at time k; the formula for calculating the pass rate Qr is: ,in This is the result of the judgment.

[0098] It should be noted that the prediction error at time k is:

[0099] When ((|Pk^{text{pre}}andPk^{text{actual}}| / Pk^{text{actual}}))≤5%, \(Bk=1) (qualified as qualified), otherwise \(Bk=0) (qualified as unqualified).

[0100] As an optional embodiment, the main terminal is also used to predict output correction, generating correction coefficients based on maintenance plans, power curtailment plans, equipment status, and installed capacity data, specifically:

[0101] The system obtains the maintenance plan of the photovoltaic power station and generates maintenance correction coefficients; it also obtains the power curtailment plan of the power grid and generates curtailment correction coefficients; and combines the power attenuation coefficient γ output by the acoustic vibration synchronous detection module to correct the prediction results using the overall correction formula, thus obtaining the corrected final predicted power.

[0102] It should be noted that the overall correction formula is:

[0103] Obtain the maintenance plan of the photovoltaic power station and generate the maintenance correction coefficient \(alpha) (\(alpha=1 and \beta) when the installed capacity of the maintenance equipment is \(beta)); obtain the power curtailment plan of the power grid and generate the curtailment correction coefficient \(delta) (\(delta=1 and \gamma{text{limit}}) when the curtailment ratio is \(gamma{text{limit}})); combine the power attenuation coefficient \(gamma) output by the acoustic vibration synchronous detection module;

[0104] By using the general correction formula:

[0105] The prediction result is corrected by \(P{text{final}}=P{text{predicted}}\times\alpha\times\delta\times(1 and\gamma)) to obtain the corrected final predicted power; where \(P{text{predicted}}) is the initial predicted power output by the photovoltaic short-term power generation prediction module.

[0106] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions that fall within the scope of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of this template.

Claims

1. A distributed photovoltaic short-term power generation prediction system based on machine learning, characterized in that: It includes a main terminal, a solar-thermal coaxial acquisition module, an acoustic-vibration synchronous detection module, and a photovoltaic short-term power generation prediction module; the solar-thermal coaxial acquisition module and the acoustic-vibration synchronous detection module are respectively communicatively connected to the main terminal, and the photovoltaic short-term power generation prediction module is also communicatively connected to the main terminal. The coaxial solar thermal acquisition module collects real-time light intensity data and photovoltaic module temperature data, and transmits them to the main terminal; the synchronous acoustic and vibration detection module collects acoustic and vibration signals of the photovoltaic equipment, identifies acoustic and vibration anomalies, and converts them into a power attenuation coefficient γ, which is then transmitted to the main terminal. After integrating the received data, the main terminal transmits it to the photovoltaic short-term power generation prediction module. The photovoltaic short-term power generation prediction module calculates the photovoltaic short-term power generation based on multi-dimensional sensor data and power attenuation coefficient γ through a hierarchical fusion prediction model, outputs the prediction results and related evaluation indicators, and sends them back to the main terminal. The photothermal coaxial acquisition module includes a visible light component and an infrared component arranged coaxially. The two are synchronized in exposure timing through a programmable logic unit and satisfy the corrected exposure time formula t3=k1×t0×(GSTC / Greal). Where k1 is the type coefficient for adapting to different photovoltaic-related equipment, t0 is the reference exposure time, GSTC is the standard light intensity, and G is the real-time light intensity; The infrared component outputs temperature data, and the visible light component outputs real-time light intensity data. The two types of data are spatially aligned using an improved ORB feature point matching technology. A curved surface projection correction step is added to address the curved surface installation characteristics of the photovoltaic module, ensuring that the temperature data and light intensity data correspond to the same module area. The solar thermal coaxial acquisition module transmits the processed light intensity data and temperature data to the main terminal, which serves as the core input parameters for short-term photovoltaic power generation prediction. The acoustic and vibration synchronous detection module includes a multi-element ultrasonic array and a triaxial vibration sensor. The multi-element ultrasonic array achieves directional focusing of ultrasonic signals through beam focusing. The focal point array spacing is designed for the partial discharge characteristics of photovoltaic inverters to improve the accuracy of partial discharge signal capture. The vibration sensor collects the vibration time-domain signal of the equipment. The time-domain signal is converted into a frequency domain spectrum by Fourier transform. After extracting specific frequency band features related to photovoltaic equipment faults, the difference is analyzed with the preset normal spectrum model. When the difference meets the set threshold, it is determined to be an acoustic vibration abnormality. The acoustic and vibration synchronous detection module transmits acoustic and vibration anomaly information to the main terminal and converts it into a power attenuation coefficient γ based on the acoustic and vibration anomaly level, which is used to correct the short-term photovoltaic power generation prediction value.

2. The distributed photovoltaic short-term power generation prediction system based on machine learning according to claim 1, characterized in that, The photovoltaic short-term power generation prediction module receives multi-dimensional sensing data transmitted from the main terminal, performs prediction calculations of photovoltaic short-term power generation based on this data, and sends the prediction results and related evaluation indicators back to the main terminal. The evaluation indicators include Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE), and their calculation formulas are as follows: ； in Let the predicted power be at time k. Let n be the actual power at time k, and n be the number of sampling points in the prediction period. ； 。 3. The distributed photovoltaic short-term power generation prediction system based on machine learning according to claim 2, characterized in that, When the photovoltaic short-term power generation prediction module performs prediction calculations, it first preprocesses the input multi-dimensional sensor data, including outlier removal based on the 3σ criterion and missing value filling based on linear interpolation. Then, a hierarchical fusion prediction model is used, which includes: Physical foundation layer: The theoretical power value is calculated based on the formula PPHYS=PSTC×(G / GSTC)×[1+αT×(Tc-TSTC)], where PSTC is the rated power under standard test conditions, G is the real-time light intensity, GSTC is the standard light intensity, αT is the power temperature coefficient, Tc is the component temperature, and TSTC is the standard temperature. Data-driven layer: A bidirectional LSTM network with an improved attention mechanism is used. The input features include time series of light intensity and temperature and historical power data. The output correction coefficient δ satisfies PML=PPHYS×δ. Dynamic fusion layer: The final prediction is achieved through the formula PFINAL=w1×PPHYS+w2×PML, where w1+w2=1; The prediction module also outputs the 95% confidence interval of the prediction result, which satisfies CI=[PFINAL and 1.96×σ,PFINAL+1.96×σ], where σ is the standard deviation of the prediction error.

4. The distributed photovoltaic short-term power generation prediction system based on machine learning according to claim 1, characterized in that, It also includes a multi-source data collaborative access unit, which is connected to the main terminal for communication. The multi-source data collaborative access unit integrates the CMS system, SCADA mirror system, marketing procurement system, business middleware, and weather forecasting system. Through the data access engine, it realizes the timed or real-time extraction of ledger data, NWP forecast weather data, real-time weather data, real-time power data, and operating status data. The multi-source data collaborative access unit performs preliminary quality screening on the collected data, removes data that obviously exceeds the physical characteristics of photovoltaic modules, and then transmits the pre-processed data to the main terminal, providing a data foundation for the model training, optimization, and evaluation of the photovoltaic short-term power generation prediction module.

5. The distributed photovoltaic short-term power generation prediction system based on machine learning according to claim 3, characterized in that, The preprocessing step also includes a hierarchical data quality closed-loop management process, specifically: The first step is data review: a three-level review mechanism of format verification, logical verification and physical verification is established. Format verification ensures that the data conforms to the preset field specifications, logical verification verifies the consistency of related data, and physical verification removes data that exceeds the limit based on the rated parameters of photovoltaic modules. The second step is anomaly classification and source tracing: the abnormal data is classified into three categories: missing, mutation, and deviation. The cause of the anomaly is located through data source tracing technology, and a data anomaly source tracing report is generated. The third step is differentiated repair: customized repair strategies are adopted for different types of anomalies. Missing data are interpolated using historical similar meteorological day data, abrupt data are smoothed using a sliding window and corrected for trends in adjacent time periods, and deviation data are corrected inversely based on the meteorological and power Pearson correlation coefficients to ensure that the repaired data matches the actual meteorological conditions. The fourth step is quality feedback: calculate the data pass rate and correlate it with the prediction accuracy to form a closed-loop feedback between data quality and prediction effect, and dynamically optimize the review and remediation strategies.

6. The distributed photovoltaic short-term power generation prediction system based on machine learning according to claim 3, characterized in that: The photovoltaic short-term power generation prediction module is designed for distributed photovoltaic scenarios, featuring a dynamic reference station cluster and a multi-factor expansion and fusion prediction scheme. The specific process is as follows: The first step is to dynamically construct the benchmark station cluster: using geographical location, climate similarity, and data qualification rate as screening criteria, distributed photovoltaic power stations in the region are selected in real time to form a benchmark station cluster. The cluster members are updated regularly based on the latest meteorological data and equipment status, and faulty power stations are removed and those under maintenance are repaired. The second step is cluster-level model training: taking the total actual power of the base station cluster as the prediction target, the input is regional NWP prediction meteorological data, cluster historical power data and real-time light intensity and temperature data of the photothermal coaxial acquisition module. The improved attention mechanism bidirectional LSTM and XGBoost integrated model are used to construct the meteorological and power mapping relationship and output the cluster-level short-term predicted power. The third step is multi-factor capacity expansion calculation: For distributed photovoltaic power stations with incomplete data in the region, meteorological similarity factor, equipment aging factor, and installed capacity factor are introduced to build an expansion model, and the predicted power is calculated through the expansion model. The fourth step is result correction: the predicted power obtained from the capacity expansion calculation is corrected by combining the power attenuation coefficient γ of the acoustic vibration synchronization detection module to obtain the final predicted power.

7. A distributed photovoltaic short-term power generation prediction system based on machine learning according to claim 2, characterized in that, Evaluation indicators also include correlation coefficient r, accuracy Cr, and pass rate Qr; the formula for calculating the correlation coefficient r is: r= , where P1 is the average actual power during the statistical period, and P2 is the average predicted power during the statistical period; The formula for calculating accuracy Cr is: ,in Let be the photovoltaic equipment's operating capacity at time k; the formula for calculating the pass rate Qr is: ,in This is the result of the judgment.

8. The distributed photovoltaic short-term power generation prediction system based on machine learning according to claim 1, characterized in that, The main terminal is also used for predictive output correction, generating correction coefficients based on maintenance plans, power curtailment plans, equipment status, and installed capacity data. Specifically: Obtain the maintenance plan of the photovoltaic power station and generate maintenance correction coefficients; The power grid curtailment plan is obtained, and a curtailment correction coefficient is generated. Combined with the power attenuation coefficient γ output by the acoustic vibration synchronous detection module, the prediction results are corrected using the overall correction formula to obtain the final predicted power after correction.

Images