Photovoltaic inverter power prediction method

Abstract

The invention discloses a photovoltaic inverter power prediction method. The method comprises the following steps: acquiring an inverter sweep point set; based on the inverter sweep point set, sparse I-V shape features representing a non-uniform state are generated; calculating a bypass risk index based on the sparse I-V shape features; when the bypass risk index is higher than a preset risk threshold value, according to the plurality of local power peak values, performing weighting calculation to obtain expected power; obtaining baseline power prediction; and based on the bypass risk index, fusing the expected power with baseline power prediction to generate a short-term prediction curve. According to the method, the observation capability of the non-uniform working condition is improved by utilizing the endogenous scanning point data of the inverter, the prediction target is corrected by solving the multi-peak expected power, and the power prediction accuracy in complex scenes such as morning and night, local shielding and the like is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of photovoltaic inverters, and particularly relates to a photovoltaic inverter power prediction method. BACKGROUND

[0002] As a key component of renewable energy, photovoltaic power generation has inherent intermittency and volatility, which poses a challenge to the stable operation of the power grid. Therefore, accurate and reliable short-time power prediction of photovoltaic inverters is not only the basis for ensuring the safe and stable dispatching of the power system, but also the core technical prerequisite for optimizing power generation plans, improving power market transaction efficiency, and maximizing the economic benefits of photovoltaic power stations, and has important research and application value.

[0003] Currently, the mainstream photovoltaic power prediction technology path mainly relies on meteorological information and massive historical power data. These methods usually use statistical tools such as time series analysis and regression models, or use machine learning algorithms such as neural networks and support vector machines, to build a mapping model from photovoltaic plane irradiance, environmental temperature, wind speed and other meteorological inputs to inverter output power. Some schemes also combine a simplified photovoltaic component physical model to estimate the ideal output power under specific working conditions, and use the estimated value as an input feature of the model.

[0004] However, the existing prediction methods often have a significant decrease in prediction accuracy when facing complex non-uniform working conditions such as morning contamination superimposed with dew and dynamic local shading. This is mainly due to the fundamental limitations of existing technologies in system state observation and prediction target setting. Specifically, traditional methods generally rely on minute-level or even lower-resolution inverter steady-state telemetry data, which is insufficient in data dimension and sampling rate to reveal the complex multi-peak morphology of the power-voltage (P-V) curve caused by non-uniform effects. The observation defect of this "invisible" problem makes it impossible to accurately quantify key states such as the early activation of bypass diodes and the evolution degree of non-uniform contamination in real time. Furthermore, due to the inability to reliably observe the multi-peak phenomenon, existing prediction models are still based on the basic assumption that there is always a single global maximum power point (MPP) in the photovoltaic array. When the actual P-V curve presents multiple peaks, there is a fundamental mismatch between the "single peak" target of the prediction model and the "multi-peak" physical reality of the photovoltaic array, resulting in a large deviation between the power value predicted by the model and the local peak value actually locked by the inverter MPPT algorithm, thus causing systematic prediction errors. SUMMARY

[0005] The application provides a photovoltaic inverter power prediction method.

[0006] Technical scheme: According to one aspect of the application, a photovoltaic inverter power prediction method comprises: obtaining an inverter scanning point set; generating a sparse I-V shape feature representing a non-uniform state based on the inverter scanning point set; calculating a bypass risk index based on the sparse I-V shape feature; when the bypass risk index is higher than a preset risk threshold, calculating an expected power according to multiple local power peaks by weighting; obtaining a baseline power prediction; fusing the expected power and the baseline power prediction to generate a short-time prediction curve based on the bypass risk index.

[0007] Beneficial effects: the application can cope with non-uniform conditions such as morning-evening contamination and local shielding, avoiding systematic overestimation or underestimation caused by the inability to perceive multiple peaks in traditional methods; by solving the expected power, a more reasonable prediction value in physics and probability is given, reducing the risk of the MPPT algorithm falling into a local peak; the accuracy and robustness of short-time power prediction are improved, providing reliable data support for grid-friendly dispatching and power station fine operation. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 is the overall flowchart of the application.

[0009] Figure 2 is the flowchart of the application embodiment for generating a sparse I-V shape feature representing a non-uniform state.

[0010] Figure 3 is the flowchart of the application embodiment for calculating the step degree.

[0011] Figure 4 is the flowchart of the application embodiment for calculating a bypass risk index.

[0012] Figure 5 is a flowchart of an application embodiment for determining the source of a non-uniform state.

[0013] Figure 6 is the flowchart of the application embodiment for fusion based on the bypass risk index.

[0014] Figure 7 is the flowchart of the application embodiment for initializing the daily reference coverage degree within the start-up window period of the photovoltaic system. DETAILED DESCRIPTION

[0015] In order to better understand the technical scheme of the present application, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work should fall within the scope of the present application.

[0016] Embodiment one: overall framework of photovoltaic inverter power prediction The present embodiment generally describes a framework of a photovoltaic inverter power prediction method, which is particularly suitable for solving the problem of power prediction accuracy of a photovoltaic system under complex working conditions such as dawn and dusk, non-uniform soiling, etc.

[0017] Specifically, a photovoltaic inverter power prediction method comprises the following steps: Step S101, obtaining an inverter scanning point set.

[0018] In the present embodiment, the inverter scanning point set does not refer to the 1-minute or 5-minute interval steady-state telemetry data uploaded by the inverter conventionally, but refers to a series of discrete voltage-power sampling point pairs generated by voltage scanning performed by the inverter in the process of performing maximum power point tracking (MPPT) to find the global optimal operating point. These scanning point data are usually recorded internally by the inverter, but are not utilized by traditional power prediction methods. In the present embodiment, through specific data acquisition interfaces or protocol analysis, these high-value scanning point data which can directly reflect the output characteristics of photovoltaic modules can be actively obtained. The traditional steady-state measurement points cannot reveal the complete shape of the P-V curve, especially when multiple peaks appear, while the use of the scanning point set provides the most original and direct data basis for subsequent shape analysis.

[0019] Step S102, generating sparse I-V shape features representing non-uniform states based on the inverter scanning point set.

[0020] In the present embodiment, the discrete and sparse scanning point data are converted into a set of structured features which can stably and quantitatively describe the shape of the P-V or I-V curve, i.e. electrical fingerprints, which can directly reflect the non-uniform effects caused by sand, dew, local shading, etc. Specifically, the sparse I-V shape features can include but are not limited to: step degree for representing the power step introduced by the bypass diode triggering, multi-peak degree for representing the intensity of the multi-peak structure of the P-V curve, voltage drift of the maximum power point relative to the clean state, and platform degree reflecting the flatness of the curve platform region, etc. By constructing these shape features, the present embodiment changes the prediction problem from simply relying on the numerical comparison of energy or current to pattern recognition of geometric shape, thereby realizing effective observation of early or mild non-uniform soiling under low-resolution telemetry data.

[0021] Step S103, based on the sparse I-V shape feature, calculate the bypass risk index, and determine the expected power, which is calculated according to the weighted sum of multiple local power peaks when the bypass risk index is higher than the preset risk threshold.

[0022] In a preferred embodiment, a dual-path power solving mechanism is established. Using the shape feature obtained in the previous step and external operating condition data, the bypass risk index R_bp is calculated, which is used to quantify the risk of power loss caused by the MPPT algorithm of the inverter tracking to a local optimal peak under the current operating condition. When the bypass risk index is lower than the preset threshold, it indicates that the system is working in a single-peak or approximately single-peak stable state, and the expected power at this time can be directly taken as the global maximum power point on the P-V curve. However, when the risk index is higher than the threshold, a new solving path is started: instead of finding a single maximum power point, all local power peaks on the P-V curve are identified and their probability weighted sum is calculated to obtain an expected power P_hat in the mathematical expectation sense, which quantifies the uncertainty brought by multiple peaks. The target is no longer the most likely peak, but the weighted average of all possible peaks, thereby changing the traditional prediction objective function to cope with the bypass risk.

[0023] Step S104, obtain the baseline power prediction.

[0024] Exemplarily, the baseline power prediction can come from any existing or conventional power prediction model, which can be a statistical model based on historical time series data (such as ARIMA, LSTM), or a hybrid model based on weather forecast and photovoltaic physical model, which is coupled with the existing prediction system as an enhancement module rather than a complete replacement. Among them, the obtained baseline power prediction P_base is regarded as a reference prediction that performs well under normal operating conditions but may have deviations under non-uniform scenarios.

[0025] Step S105, based on the bypass risk index, fuse the expected power with the baseline power prediction to generate a short-term prediction curve.

[0026] The embodiment dynamically weights and fuses the expected power P_hat and the baseline prediction P_base through a risk-aware fusion strategy. The weight coefficient of the fusion is a monotonically increasing function of the bypass risk index R_bp. When the bypass risk is higher, the final prediction result is more inclined to believe the expected power calculated by the embodiment; when the risk is very low, it mainly depends on the traditional baseline prediction. Through the technical solution of the embodiment, targeted correction of the traditional model is realized, intervention is only performed in specific periods of time (such as uneven soiling during dawn and dusk) where high risk is identified, compatibility with the original system is maintained during the remaining stable periods, and finally a short-term prediction curve P_forecast_60m that is stable and scene-adaptive is generated.

[0027] Embodiment two: sparse I-V shape feature generation method based on inverter scanning points The embodiment describes how to convert original discrete scanning points with working condition noise into standardized and analyzable sparse I-V fitting curves.

[0028] In a preferred embodiment, before constructing the sparse I-V fitting curve, the voltage and power discrete points in the inverter scanning point set are further mapped to the preset reference temperature and reference irradiance conditions according to the component temperature and plane irradiance to generate temperature-irradiance-consistent scanning points for subsequent shape comparison. Specifically, this step is called temperature-irradiance-consistent preprocessing, which eliminates the stretching and drifting of I-V / P-V curves caused by different time points and different working conditions (mainly temperature T_module and plane irradiance POA), so that the subsequent shape analysis can focus on the morphological distortion caused by non-uniform effects.

[0029] Optionally, using the temperature coefficient provided by the photovoltaic module manufacturer, each "voltage U, power P" point in the scanning point set is corrected to the corresponding value at the reference temperature T_ref (for example, 25 degrees Celsius) according to the current measured component temperature T_module. Similarly, all scanning points within a short time window (for example, 5 minutes) can be normalized according to the irradiance, assuming that the power P and the plane irradiance POA are approximately linearly related. Through the technical solution of the embodiment, a set of temperature-irradiance-consistent scanning points_S_TI that eliminates the differences in working conditions is obtained, laying a foundation for constructing pure shape curves.

[0030] Further, the step of generating the sparse I-V shape feature representing the non-uniform state comprises: Step S201, constructing a sparse I-V fitting curve according to the inverter scanning point set.

[0031] In a specific embodiment, the step of constructing a sparse I-V fitting curve comprises: The inverter scanning point set is fitted as a sparse I-V piecewise curve, the sparse I-V piecewise curve includes an open circuit section, a platform section and a steep drop section; and a power step is identified in the fitting residual of the platform section, so as to construct the sparse I-V fitting curve.

[0032] Preferably, instead of using a single complex function to fit all scanning points, a divide-and-conquer strategy is adopted. The pre-processed temperature irradiation uniform scanning points_S_TI are sorted by voltage U. In the high voltage U interval, an open circuit section is identified, the power P changes gently with the voltage U, and a low-order polynomial (such as first or second order) can be used for fitting.

[0033] Further, in the medium voltage U interval, a platform section is identified, i.e. the region where the maximum power point is located, the power P value is relatively high and does not change much, and a constant value plus a perturbation term can be used to represent it.

[0034] Further, in the low voltage U interval, a steep drop section is identified, the power P decreases rapidly with the decrease of the voltage U, and a monotonically decreasing function (such as an exponential decay function) can be used for fitting.

[0035] Further, the three section fitting curves are connected to form a sparse I-V piecewise curve_IV_piece.

[0036] According to a further improvement of the present application, the fitting residual of the platform section is analyzed. The conduction of the bypass diode will cause a small, step-like power drop on the originally flat power platform. By calculating the discrete gradient of the platform section power P, or using a sliding difference method, these power mutation points, i.e. power steps, can be detected.

[0037] Further, the identified power step information is superimposed on the piecewise fitting curve to form the final sparse I-V fitting curve_IV_sparse which can finely reflect the non-uniform effect.

[0038] In step S202, the sparse I-V fitting curve is analyzed to calculate the sparse I-V shape features, the sparse I-V shape features include at least one feature selected from the following group: step degree, multimodality, maximum power point voltage drift and platform degree.

[0039] In one specific embodiment, after obtaining the structured sparse I-V fitting curve, a series of quantitative indicators can be extracted therefrom. Exemplarily, the number, amplitude and width of the power steps can be extracted from the plateau segment of the curve for calculating the step degree; all local power peaks (including main peaks and secondary peaks) can be identified from the entire curve for calculating the multi-peak degree; the maximum power point voltage of the current curve can be compared with the reference curve of the historical cleaning day to obtain the maximum power point voltage drift; and the voltage width of the plateau segment can be calculated as a proportion of the total open-circuit voltage to obtain the plateau degree. These features will be described in detail in subsequent embodiments.

[0040] Embodiment Three: Specific Quantitative Calculation Method of Sparse I-V Shape Features This embodiment describes in detail how to accurately calculate two core shape features, step degree and multi-peak degree, from the constructed sparse I-V fitting curve, so as to convert the visual curve shape difference into stable numerical indicators that can be input into the downstream model.

[0041] Calculation of Step Degree: In this embodiment, the calculation of step degree includes: Step S301: determining the sum of power step amplitudes from the plateau segment of the sparse I-V fitting curve, and dividing by the reference power obtained from the cleaning baseline curve to obtain a power ratio. The cleaning baseline curve IV_clean_ref refers to the standard I-V output curve of the photovoltaic module under ideal (i.e., without any non-uniform pollution or shading) working conditions. This curve can be obtained in various ways. Exemplarily, it can be obtained by field calibration test at the initial operation of the photovoltaic power station or after cleaning of the module; or it can be obtained by using the data provided by the manufacturer under standard test conditions (STC) and correcting it according to the real-time working conditions. It is the "golden standard" or reference system for all subsequent shape distortion analyses. The reference power P_plateau_clean specifically refers to the average power value of the plateau segment of the cleaning baseline curve in the same voltage interval as the current curve to be tested. The power step amplitude ΔP_step refers to the height of each power step identified in Embodiment Two, i.e., the power drop amplitude. If there are multiple steps in the plateau segment, their amplitudes are added to obtain the sum of power step amplitudes ΣΔP_step,i. Through the technical solution of this embodiment, the total power step absolute height is normalized by the plateau power under ideal conditions to obtain a dimensionless power ratio that represents the relative severity of power drop.

[0042] Step S302, determine the plateau length from the plateau section of the sparse I-V fitting curve, and divide by the open circuit voltage obtained from the clean baseline curve to obtain a voltage ratio. Wherein, the plateau length L_plateau refers to the difference between the start voltage and the end voltage of the plateau section identified on the sparse I-V fitting curve, which reflects the width of the flat region near the maximum power point. The open circuit voltage U_oc refers to the open circuit voltage value obtained from the clean baseline curve. Through the technical scheme of the embodiment, the absolute voltage width of the plateau region is normalized by the open circuit voltage in the ideal state to obtain a dimensionless voltage ratio representing the relative width of the plateau region.

[0043] Step S303, multiply the power ratio by the voltage ratio to obtain the step degree. Specifically, the calculation formula of the step degree_step_metric can be represented as: i (ΔP_step,i) / P_plateau_clean)×(L_plateau / U_oc); Wherein, the step degree_step_metric is a dimensionless value, the larger the value, the more serious the power loss caused by the conduction of the bypass diode, and the wider the voltage range affected. Exemplarily, a small step that only affects a very narrow voltage range will have a small step degree; while a deep step that spans most of the plateau region will significantly increase the step degree. Through the technical scheme of the embodiment, the depth and width information of the step in two dimensions is integrated into one index, which can more comprehensively reflect the severity of the bypass effect.

[0044] In some optional embodiments, if the accurate clean baseline curve cannot be obtained, the reference power P_plateau_clean can also be approximately replaced by the power value P_peak,max of the main peak (i.e. the global maximum power peak) of the current sparse I-V fitting curve. Correspondingly, the open circuit voltage U_oc can also be replaced by the open circuit voltage value obtained by the current curve fitting. Although the accuracy of this replacement scheme is slightly reduced, it provides a feasible and self-consistent calculation path when baseline data is lacking.

[0045] Calculation of multimodality: In this embodiment, the calculation of multimodality includes: Step S304, identify the local peak list from the sparse I-V fitting curve, determine the main peak power. By taking the derivative of the sparse I-V fitting curve (more precisely, its corresponding P-V curve), all points with zero first derivative and negative second derivative can be found, which constitute the local peak list P_peaks, containing the voltage U_peak_k and power P_peak_k information of each local peak. The main peak power P_peak,max refers to the power of the peak with the largest power value in the list.

[0046] Step S305, for any local peak other than the main peak, calculate its relative power difference with the main peak power, determine its voltage interval with the adjacent peak. The relative power difference ΔP_peak,k is the difference between the main peak power P_peak,max and the power P_peak_k of the kth local peak (non-main peak), which reflects the degree of power loss of the secondary peak compared to the global optimum. The voltage interval ΔU_peak,k is the voltage difference between the kth peak and its adjacent previous peak in the voltage-ordered local peak list, which reflects the degree of separation of each local optimum on the voltage axis.

[0047] Step S306, multiply the relative power difference by the open-circuit voltage normalized voltage interval to obtain the product term of each local peak.

[0048] Step S307, sum the product terms of all local peaks, normalize the sum value by the main peak power to obtain the multimodal metric. Specifically, the calculation formula of the multimodal metric_multimodal_metric can be expressed as: _multimodal_metric=(1 / (P_peak,max))∑ N_peaks k=2 (ΔP_peak,k×(ΔU_peak,k) / U_oc); where the multimodal metric_multimodal_metric is also a dimensionless value. It not only considers the existence of secondary peaks (N_peaks>1), but also weighs the harmfulness (reflected by ΔP_peak,k, the larger the difference, the more harmful) and the misleadingness (reflected by ΔU_peak,k, the larger the interval, the more likely to mislead the MPPT algorithm) of the secondary peaks. A P-V curve with a high multimodal metric means that its energy terrain is very complex, and conventional MPPT algorithms are prone to fall into local optima, resulting in power generation loss. This embodiment provides a direct quantitative basis for the downstream risk judgment.

[0049] Embodiment four: calculation of bypass risk index and solution method of expected power This embodiment discloses the complete logical chain from identifying risks to changing the prediction target. In this embodiment, the steps for calculating the bypass risk index are described, which specifically include: Step S401, according to sparse I-V shape characteristics and working condition data including incident angle and bypass threshold prediction, a set of sub-item risk scores are determined. Among them, the sparse I-V shape characteristics mainly include the step degree_step_metric and multimodal degree_multimodal_metric calculated in embodiment three, etc. The working condition data includes but is not limited to: solar incident angle AOI, bypass threshold prediction_I_thr, and optional dew indication_dew_index, etc. Further, each of the above input features is converted into a normalized sub-item risk score by an independent mapping function. Exemplarily, the sub-item risk score r_step = f1(step degree_step_metric); the sub-item risk score r_multi = f2(multimodal degree_multimodal_metric); the sub-item risk score r_aoi = f3(incident angle AOI); wherein the functions f1, f2, f3 can be simple linear functions, piecewise functions or lookup tables, which uniformly map inputs of different physical units and dimensions to risk scores in the interval [0, 1]; usually designed to be monotonic, for example, the greater the step degree, the higher the r_step score.

[0050] Step S402, a set of sub-item risk scores are weighted and combined to generate a bypass risk index. The sub-item risk scores obtained in the above step are combined into the final bypass risk index R_bp by linear weighting: R_bp = w1 x r_step + w2 x r_multi + w3 x r_shift + w4 x r_aoi + w5 x r_dew + w6 x r_margin; wherein w1, w2, w3... are the weight coefficients of each sub-item risk, and their sum is 1. These weight coefficients can be pre-set by expert experience, or preferably, trained and optimized on a large amount of historical data by machine learning method. The final R_bp is a value between 0 and 1, which intuitively represents the risk probability of current bypass effect and causes MPPT mismatch.

[0051] r_shift is a sub-risk score related to the maximum power point voltage shift, which is calculated from the shape feature voltage shift_dV_mpp_AOI, and is used to quantify the degree of deviation of the current maximum power point from the cleaning baseline. Voltage shift is another important indicator of mismatch within the photovoltaic array; r_dew is a sub-risk score related to dew indication, and dew indication_dew_index is an external input, which can preferably be calculated by combining meteorological data (such as temperature, humidity, wind speed) with a specific model, or directly measured by a field sensor. Morning dew can exacerbate the accumulation and adhesion of dust at the bottom of the module, and is an important environmental factor that induces high risk of morning bypass; r_margin is a sub-risk score related to bypass safety margin, which integrates the bypass threshold prediction_I_thr and the string current margin I_margin, and is used to evaluate the safety space of the current photovoltaic string working current from the threshold value of triggering the bypass diode conduction. The smaller the margin, the higher the risk; the weight coefficients w1 to w6 are weight coefficients corresponding to the six sub-risk scores respectively. These weight coefficients are positive values, and their sum is 1 (w1 + w2 +... + w6 = 1). These weight values reflect the importance of different risk factors in the comprehensive evaluation, which can be calibrated by expert experience, or optimized by machine learning algorithm on a large amount of historical data.

[0052] According to the further improvement of the present application, the step of determining the expected power is described, which specifically includes: Step S403, when the bypass risk index exceeds the preset risk threshold, the plurality of local power peaks obtained from the sparse I-V shape feature are weighted and summed to determine the expected power. This is a key decision gating link. Illustratively, a risk threshold_tau_bp (for example, its value range can be between 0.3 and 0.7, which is debugged according to the specific situation of the power station) is preset; compare R_bp calculated in step S402 with_tau_bp; if R_bp<_tau_bp, the system judges that the current risk is low, adopts the conventional prediction logic, and the expected power is the global unique maximum power point on the sparse I-V fitting curve; if R_bp≥_tau_bp, the system judges that the risk is high, and no longer relies on a single maximum power point, and instead executes a new expected power solving logic.

[0053] Further, the plurality of local power peaks are weighted and summed, including: Step S404, for each local power peak, apply a softmax function to the weight input vector to generate a multi-peak weight. First, for each local power peak (denoted as the kth peak), a weight input vector z_k needs to be constructed. In a preferred embodiment, the step of generating a weight input vector for each local power peak includes: applying a set of predefined weight coefficients to weight sum the step degree and the multi-peak degree, the non-uniform severity, the incident angle, and the local peak curvature of the local power peak selected from the sparse I-V shape feature, to obtain the weight input vector; the formula can be expressed as: z_k = a_0 + a_1 × step degree + a_2 × multi-peak degree + a_3 × severity + a_4 × AOI + a_5 × peak curvature_curvk; wherein a_0 to a_5 are internal parameters of the model, which can be obtained by learning; the non-uniform severity is an index for evaluating the non-uniform state; the peak curvature_curv_k is the second derivative near the kth peak point on the P-V curve, reflecting the sharpness of the peak, and generally the more sharp peaks have higher weights. Further, the obtained z_k vector (each peak has a z value) is transformed by a softmax function to obtain a normalized multi-peak weight w_k: w_k = exp(z_k) / ∑_j exp(z_j).

[0054] Preferably, the application of the softmax function can convert the z_k vector of any real number value into a probability distribution with a sum of 1, where each w_k can be understood as the probability of the inverter actually working at the kth local peak.

[0055] Step S405, use the multi-peak weight to weight sum the plurality of local power peaks. In this embodiment, the expected power P_hat is calculated by: P_hat = ∑_k w_k × (P_peak,k); wherein P_peak,k is the power value of the kth local peak. The expected power P_hat is not any actually existing power peak, but a mathematical expectation after considering all possibilities. It integrates the bypass uncertainty in an inherent and feedforward manner into the prediction value itself, so that the prediction result naturally contains the consideration of risk, which has higher timeliness and accuracy compared with the traditional method of risk warning or correction after prediction.

[0056] Embodiment five: joint discrimination method of non-uniform state source This embodiment aims to further elaborate the cause diagnosis of non-uniform state, an important link in the overall method. After identifying the non-uniform state (e.g., the bypass risk index R_bp is increased) through the methods of embodiments three and four, it is crucial to accurately distinguish whether the physical source is a transient dynamic shading (such as cloud shadow, floating object) or a relatively stable static contamination (such as sand dust, accumulated dust) for subsequent adoption of what kind of correction strategy (e.g., whether to update the daily contamination benchmark). This embodiment proposes a discriminant method based on the joint triggering of multi-dimensional electrical characteristics.

[0057] In a specific embodiment, a step for determining the source of non-uniform state, the step comprising: Step S501, evaluating the frequency domain stability of the power sequence. Different physical causes of non-uniform state will leave different time fingerprints on the output power time sequence of the inverter. Specifically, the step of evaluating the frequency domain stability of the power sequence comprises: Calculating the high-frequency flicker index and the low-frequency smoothing index of the power sequence. The power sequence refers to the time sequence data of the continuous power measured value P_meas obtained from the inverter steady-state measurement point within a recent time window (e.g., 5 to 15 minutes).

[0058] Further, the high-frequency flicker index HFI is used to quantify the rapid and severe fluctuations of power. In a specific implementation, a band-pass filter can be applied to the power sequence first (e.g., retaining frequency components between 0.2 Hz and 2 Hz), and then the variance or standard deviation of the filtered signal is calculated, which is the HFI. Physically, dynamic shading caused by moving clouds or wind-blown leaves, etc. usually has clear edges, and when it sweeps across the surface of the component, it will cause a sharp drop and recovery of power, thereby generating significant high-frequency components and corresponding higher HFI values.

[0059] Further, the low-frequency smoothing index LFI is used to quantify the slow and smooth changes of power. In a specific implementation, a low-pass filter can be applied to the power sequence first (e.g., only retaining frequency components below 0.05 Hz), and then the reciprocal of the standard deviation of the filtered signal is calculated and normalized to obtain LFI. Physically, the power attenuation caused by sand dust or accumulated dust is relatively stable or changes slowly on a minute scale (mainly changes with the sun angle), so its power sequence shows a smooth low-frequency trend, corresponding to a higher LFI value.

[0060] Step S502, verifying the monotonicity of power loss change with incident angle. Specifically, the step of verifying the monotonicity of power loss change with incident angle comprises: The time series of power loss and incident angle are subjected to statistical correlation coefficient test. The power loss can be defined as the difference between the baseline predicted power P_base and the real-time measured power P_meas. The incident angle AOI is the angle between the sunlight and the normal of the PV module plane, which can be calculated from the time and geographical location information.

[0061] In one specific implementation, the time series of power loss and incident angle AOI are collected synchronously within a relatively long time window (e.g. 30 to 60 minutes). Further, the Pearson correlation coefficient between the two time series is calculated, and when the correlation coefficient is a significant positive value (e.g. greater than a preset threshold, such as 0.6), it is determined that the monotonicity condition is satisfied. The physical principle is that when the incident angle AOI increases (e.g. during the dawn and dusk period), the sunlight needs to penetrate a thicker layer of dust, resulting in a sharp increase in optical loss, so the power loss and AOI show strong positive correlation, while the power loss caused by dynamic shading is mainly the geometric shading effect, which has no direct and monotonic association with the incident angle.

[0062] Step S503, evaluate the consistency of the power loss with the daily reference coverage. The prior information on the daily scale is introduced as a reference, where the daily reference coverage_dust_cov_day is a state variable maintained by the method, representing the overall contamination level of the day, which changes very slowly within the day.

[0063] In one specific implementation, the relative size of the power loss calculated in the current short time window is compared with the value of_dust_cov_day. If the current power loss rate and the value of_dust_cov_day remain consistent within a reasonable error range (e.g. ±5%), it is determined to be "consistent". The_dust_cov_day represents the system's expectation of the contamination of the day, and the power loss caused by the contamination itself should be consistent with this expectation. If a power loss event far exceeds the current reference (e.g. a sudden 30% power loss on a very clean day), the event is likely to be caused by external factors unrelated to the reference, i.e. shading.

[0064] Step S504, jointly determine whether the source of the non-uniform state is shading or dust according to the evaluation results of the frequency domain stability, the test results of the monotonicity, and the evaluation results of the consistency. This embodiment is a multi-condition fusion decision process, which uses a rigid trigger logic to improve the accuracy and robustness of the discrimination.

[0065] Exemplarily, the rule for determining "shading" is: when and only when the HFI calculated in step S501 is significantly higher than the threshold, and the AOI monotonicity in step S502 does not satisfy, and the reference consistency in step S503 does not satisfy, it is jointly determined that the source is "shading".

[0066] For example, the rule for determining "sand dust" is: the LFI calculated in step S501 is significantly higher than the threshold value, the AOI monotonicity in step S502 is satisfied, and the reference consistency in step S503 is satisfied, then the source is determined to be "sand dust".

[0067] For example, if any one of the above complete condition combinations is not satisfied, the source can be determined to be "uncertain", and a conservative strategy is adopted. Such a multi-evidence joint decision mechanism avoids misjudgment of a single indicator due to noise or boundary conditions, and constitutes an important innovation of the present application.

[0068] Alternatively, as an alternative to rigid rules, a scoring system can also be designed, and whether each condition is satisfied or not will increase or decrease the score for the "occlusion" or "sand dust" category, and finally the total score is used for determination. Alternatively, a small decision tree or naive Bayes classifier can be trained using labeled historical data to perform this fusion discrimination task. Further, in order to make the discrimination result more stable, the discrimination results of continuous multiple time windows can be majority voted, or a hysteresis threshold can be set to avoid frequent switching of results in critical states.

[0069] Embodiment six: risk-aware adaptive prediction fusion strategy Instead of simply replacing the baseline prediction P_base with the expected power P_hat, this embodiment designs an intelligent fusion mechanism so that this replacement is smooth, adaptive and robust.

[0070] In a specific implementation, the step of fusing based on the bypass risk index includes the following process of determining the fusion coefficient: Step S601, the bypass risk index is mapped through a monotonically increasing function to obtain an initial fusion coefficient. In this embodiment, the bypass risk index R_bp is calculated by embodiment four, which is converted into a weight coefficient directly used for fusion, and the value range is [0, 1]. For example, the monotonically increasing function g_mono can have various forms. In a simple implementation, it can be a linear function, for example: the initial fusion coefficient alpha0 = R_bp. In another preferred embodiment, it can be a piecewise linear function to achieve different sensitivities in different risk intervals. For example, when R_bp is lower than a certain low threshold (such as 0.2), alpha0 is 0; when R_bp is higher than a certain high threshold (such as 0.8), alpha0 is a certain upper limit value; between the two thresholds, alpha0 increases linearly. Such a design can achieve the clear logic of completely relying on the baseline in low risk and mainly relying on the expected power in high risk.

[0071] Step S602, when the consistency flag for characterizing the expected power stability is false, a preset upper limit value is used to limit the initial fusion coefficient to generate the final fusion coefficient for fusion. The embodiment introduces an important safety or anti-shake mechanism. In a specific implementation, the generation of the consistency flag_consistency_flag includes: The fluctuation of the expected power in adjacent time windows is evaluated. Specifically, the change rate of the newly calculated expected power P_hat compared to P_hat of the previous time window can be calculated, and if the change rate exceeds the preset fluctuation threshold (for example, 10% / minute), it is considered that the fluctuation is large.

[0072] Further, the stability of the effective peak list corresponding to the plurality of local power peaks in adjacent time windows is checked. Specifically, the effective peak list P_peaks_eff of the current time window and the previous time window can be compared. If the number of peaks, the relative position or power order of each peak changes dramatically, it indicates that the peak list is unstable.

[0073] Further, according to the fluctuation evaluation result of the expected power and the stability check result of the effective peak list, the true or false of the consistency flag is determined together. When P_hat fluctuation is not large and the effective peak list remains stable, the_consistency_flag is set to true, otherwise it is false.

[0074] According to a further improvement of the present application, when_consistency_flag is false, it means that although the risk index is high, in some embodiments, the calculated expected power P_hat can be in a transient state of instability, at this time it is not appropriate to rely on it too much, and an upper limit value alpha_cap (for example, the public range is 0.5-0.8) needs to be applied to the obtained initial fusion coefficient alpha0 to limit it, that is, the final fusion coefficient alpha_Rbp_final = min(alpha0, alpha_cap). Through the technical solution of the embodiment, the robustness of the fusion process is improved, and the severe disturbance to the final prediction curve caused by the transient instability in the model is avoided.

[0075] Further, the short-term forecast curve P_forecast_60m is generated by the following linear fusion formula: P_forecast_60m=α_Rbp_final×P_hat_final+(1-α_Rbp_final)×P_base; Where P_hat_final is the expected power obtained in Embodiment Four (which may have been post-processed), and P_base is the baseline power prediction. This fusion strategy is risk-driven and constrained by stability, embodying intelligence and engineering practicality.

[0076] The embodiment details the full life cycle management method of the key state variable daily reference coverage _dust_cov_day. The variable is the bridge connecting the fast assessment in the start-up period and the continuous tracking in the daytime, and its accuracy directly affects the accuracy of the non-uniform source discrimination and the overall prediction.

[0077] In a specific implementation, the method further includes a step of initializing the daily reference coverage in the start-up window period of the photovoltaic system, which includes: Step S701, determining an initial estimate value according to the inverter scanning point set. Illustratively, the start-up window period specifically refers to the initial stage of the photovoltaic system from standby to start grid-connected power generation in the morning of each day (for example, within 5 to 15 minutes after sunrise), at which time the irradiance is low and changes rapidly. Further, a temporary and rough sparse I-V shape is quickly constructed by using the extremely limited inverter scanning point set_S generated by the inverter start-up self-check or the first MPPT scanning. Further, by analyzing the distortion degree (for example, deviation from the cleaning baseline) of the initial shape, an initial estimate value about the daily pollution condition can be obtained.

[0078] Step S702, compensating the initial estimate value according to the change rate of the component temperature to generate the daily reference coverage for use in subsequent steps. Illustratively, in the start-up window period, the component temperature T_module will rapidly rise from the ambient temperature, and the temperature rise itself will cause a significant drift of the I-V / P-V curve, which effect is easy to be confused with the shape distortion caused by pollution. If not distinguished, the temperature rise effect will be wrongly attributed to pollution, thereby overestimating the daily pollution degree. Preferably, the embodiment calculates the change rate dT_module_dt (i.e. the acceleration of temperature rise) of the component temperature by collecting the component temperature sequence at the adjacent time. Further, a compensation model is established to separate the shape change part caused by dT_module_dt from the obtained initial estimate value. Illustratively, the drift amount of the maximum power point voltage under different dT_module_dt can be pre-calibrated, and this part is subtracted from the total voltage drift. Further, after compensation, a daily reference coverage _dust_cov_day that more purely reflects the true pollution level is obtained for use in subsequent full-day prediction and discrimination.

[0079] In a specific implementation, the method further includes: Step S703, in the period when the bypass risk index is lower than the preset low risk threshold, the prediction residual between the short-term prediction curve and the measured power is calculated. When the bypass risk index R_bp is low enough (for example, lower than 0.2), and preferably, the consistency flag_consistency_flag is true, it indicates that the system works in a stable, predictable single-peak state. At this time, the main source of uncertainty of the model is the slight deviation of the reference coverage. In this period, the prediction residual e = P_meas - P_forecast_60m_c is calculated, where P_meas is the power uploaded by the inverter in real time, and P_forecast_60m_c is the final prediction value output.

[0080] Step S704, and according to the prediction residual, the daily reference coverage used to characterize the degree of pollution on the day is updated online to use the updated daily reference coverage for the calculation of the subsequent prediction period. A gradient descent with a limited amplitude and a small step is used to update the reference. Specifically, the update formula can be expressed as: dust_cov_day_new = dust_cov_day + _dust_cov_day × clip(e, - , );wherein dust_cov_day_new is the updated reference value; is a very small learning rate or step (for example, the public range is 0.01-0.05) to ensure the smoothness of the update process and avoid overfitting of transient disturbances; the clip(e, - , ) function is used to truncate the residual e, is a truncation threshold (for example, 2% to 8% of the rated power), its role is to ignore the residual caused by non-pollution factors (such as unidentified cloud shadows) that is too large, and further ensure the robustness of the update. Preferably, this update process should also include a weather mutation protection mechanism; that is, when a sudden change in plane irradiance POA or incident angle AOI is detected, the update of _dust_cov_day should be suspended until the weather condition returns to stable.

[0081] By implementing steps S703 and S704 in the embodiment technical solution, the online learning mechanism capable of continuously and stably self-correcting the pollution assessment in a day is realized.

[0082] This embodiment describes the last processing step from the fused prediction value to the final prediction curve published externally, which conforms to the engineering practice.

[0083] In one specific implementation, after obtaining the short-term forecast curve P_forecast_60m by the method of embodiment six, the method further comprises a step of imposing physical constraints on the short-term forecast curve, the physical constraints comprising at least one selected from the group consisting of: a non-negative constraint, an inverter rated power upper limit constraint, and a ramp rate upper limit constraint.

[0084] Specifically, these constraints are necessary guarantees to ensure that the prediction results are physically and engineering feasible.

[0085] Illustratively, the non-negative constraint is to check all power points on the forecast curve, and if there is a value less than zero, it is set to zero, because the output power of the photovoltaic inverter cannot be negative in power generation mode.

[0086] Illustratively, the inverter rated power upper limit constraint is to obtain the rated output power P_rated of the inverter. Check all power points on the forecast curve, and if there is a value greater than P_rated, limit it to P_rated, because the output power of any inverter cannot exceed the maximum capacity of its hardware design.

[0087] Illustratively, the ramp rate upper limit constraint is to obtain the maximum power change rate allowed by the inverter, i.e. the ramp rate R_max (units are usually kW / s or percentage of rated power / min), which is usually determined by grid guidelines or hardware performance. Check the power values P(t) and P(t+1) of any two adjacent time points on the forecast curve to ensure that |P(t+1) - P(t)| / Δt ≤ R_max; if not, modify P(t+1) to meet the ramp rate limit. This constraint ensures that the trend of the forecast curve is physically achievable by the inverter.

[0088] Optionally, before publishing the forecast curve externally, interface packaging and prediction footprinting will also be performed. That is, the forecast curve after physical constraints is packaged according to the data format agreed with the grid dispatch center or power trading platform (such as JSON, XML). At the same time, the current forecast curve, key intermediate variables (such as expected power P_hat, fusion coefficient alpha_Rbp_final) and input features for decision making are stored in the log or database for subsequent performance audit, fault playback and model iteration.

[0089] Embodiment nine: a specific numerical calculation case This embodiment provides a specific numerical calculation process to exemplarily illustrate the complete process of the present application from obtaining the sweep point to generating the final prediction value.

[0090] In this example, a certain PV power plant at 7:30 am, the sun incident angle AOI = 75 degrees, there is moderate non-uniform dust; inverter scanning point set_S (has been done with temperature radiation unification processing): {(U=580V, P=135kW), (U=600V, P=148kW), (U=610V, P=150kW), (U=620V, P=143kW), (U=640V, P=141kW), (U=680V, P=120kW)}; platform segment reference power P_plateau_clean = 180kW, open circuit voltage U_oc = 720V; P_base = 165kW (traditional model does not consider multi-peak, tends to overestimate); non-uniform severity = 0.6, local peak curvature information is known.

[0091] In one specific implementation, the step-by-step calculation process includes: Step A: Sparse I-V fitting and feature calculation After fitting, two local power peaks are identified from the scanning point set: the main peak P_peak1 is at (610V, 150kW), and the secondary peak P_peak2 is at (580V, 135kW); The platform segment appears a power step from 150kW to 143kW, ΔP_step = 7kW, and the platform segment voltage length L_plateau is about 50V; Calculate the step degree_step_metric = (7kW / 180kW) × (50V / 720V) ≈ 0.0027; Calculate the multi-peak degree_multimodal_metric: ΔP_peak,2 = 150-135=15kW, ΔU_peak,2 =610-580=30V; multi-peak degree ≈ (1 / 150kW) × (15kW × (30V / 720V)) = 0.0417.

[0092] Step B: Bypass risk index calculation Assuming that each sub-risk score is calculated and weighted, the bypass risk index R_bp = 0.7 is obtained.

[0093] Step C: Expected power solving Since R_bp = 0.7> threshold_tau_bp (for example, 0.5), enter the expected power solving path; Assuming that the weight coefficient a_i is known, calculate the weight input vector z_k of the two peaks: z_1 (main peak) = -0.2, z_2 (secondary peak) = 0.9; The softmax function is applied to calculate the multi-peak weight w_k: w_1 = exp(-0.2) / (exp(-0.2) + exp(0.9)) ≈ 0.819 / (0.819 + 2.46) ≈ 0.25; w_2 = exp(0.9) / (exp(-0.2) + exp(0.9)) ≈ 2.46 / (0.819 + 2.46) ≈ 0.75.

[0094] The expected power is calculated: P_hat = w_1 × P_peak1 + w_2 × P_peak2 = 0.25 × 150kW + 0.75 × 135kW = 37.5 + 101.25 = 138.75 kW.

[0095] Step D: Risk-aware fusion According to R_bp=0.7, the initial fusion coefficient alpha0=0.7 is calculated; Assuming that the consistency flag is true, no upper limit constraint is needed, and the final fusion coefficient alpha_Rbp_final = 0.7; The final forecast power is calculated: P_forecast_60m = 0.7 × P_hat + (1 - 0.7) × P_base; P_forecast_60m = 0.7 × 138.75kW + 0.3 × 165kW = 97.125 + 49.5 = 146.625 kW.

[0096] Through the technical solution of the embodiment, the final short-term forecast power is 146.625kW. This result significantly corrects the overestimation of the traditional baseline model 165kW, is more conservative than simply taking the main peak 150kW, and is more optimistic than taking the secondary peak 135kW, which is a more realistic expected value considering all possibilities.

[0097] In this embodiment, to solve the problem of insufficient observability, the application discards the dependence on low-resolution steady-state data and uses the scan point set generated by the inverter itself during the MPPT scanning process, which has been ignored in the past, as high-value input information. By fitting these discrete points, a sparse I-V fitting curve that can reflect the true output characteristics of the component is reconstructed, and a series of quantifiable shape features such as step degree and multi-peak degree are extracted from it. A new observation paradigm is established, making the multi-peak shape of the P-V curve, a key internal state, visible and measurable, thus achieving real-time and sensitive online monitoring of the non-uniform effect.

[0098] Further, for the problem of predicting target failure, the application changes the objective function of the prediction task on the basis of being able to reliably observe the multi-peak risk. When the bypass risk index exceeds the threshold, the method no longer adheres to the wrong assumption of finding a single maximum power point, but actively identifies all local power peaks on the P-V curve, calculates their respective probability weights according to their characteristics, and finally obtains an expected power in the mathematical expectation sense through weighted summation.

[0099] Through the above two-step innovation of visible and accurate calculation, the application can cope with non-uniform working conditions such as morning and evening contamination and local shading. Not only does it avoid the systematic overestimation or underestimation caused by the inability to perceive multiple peaks in traditional methods, but also by solving the expected power, it gives a more reasonable prediction value in both physics and probability, which contains the risk of the MPPT algorithm falling into a local peak. This greatly improves the accuracy and robustness of short-term power prediction, providing more reliable data support for grid-friendly dispatching and fine operation of power stations.

[0100] The preferred embodiments of the application are described in detail above, but the application is not limited to the specific details in the above embodiments. Within the technical concept of the application, various equivalent transformations of the technical solutions of the application can be made, and these equivalent transformations all belong to the protection scope of the application.

Claims

1. A photovoltaic inverter power prediction method, characterized by, The method comprises: obtaining an inverter scanning point set; generating a sparse I-V shape feature representing a non-uniform state based on the inverter scanning point set; calculating a bypass risk index based on the sparse I-V shape feature; when the bypass risk index is higher than a preset risk threshold, calculating an expected power by weighted summation of multiple local power peaks; obtaining a baseline power prediction; fusing the expected power and the baseline power prediction to generate a short-term prediction curve based on the bypass risk index.

2. The method of claim 1, wherein, The method of generating a sparse I-V shape feature representing a non-uniform state comprises: constructing a sparse I-V fitting curve based on the inverter scanning point set, including: fitting the scanning point set into a sparse I-V segmented curve containing an open circuit segment, a platform segment and a steep drop segment, and identifying a power step in the fitting residual of the platform segment; analyzing the sparse I-V fitting curve to calculate a sparse I-V shape feature; the shape feature at least includes one of: step degree, multimodality, maximum power point voltage drift and platform degree.

3. The method of claim 2, wherein, The calculation of the step degree comprises: determining the sum of power step amplitudes from the platform segment of the sparse I-V fitting curve, dividing by the reference power obtained from the clean baseline curve to obtain a power ratio; determining the platform segment voltage length from the platform segment of the sparse I-V fitting curve, dividing by the open circuit voltage obtained from the clean baseline curve to obtain a voltage ratio; multiplying the power ratio and the voltage ratio to obtain the step degree.

4. The method of claim 2, wherein, The calculation of the multimodality comprises: identifying a local peak list from the sparse I-V fitting curve, and determining the main peak power; for any local peak other than the main peak, calculating the relative power difference between the local peak and the main peak power, and determining the voltage interval between the local peak and the adjacent peak; multiplying the relative power difference and the voltage interval normalized by the open circuit voltage to obtain a product term of each local peak; summing the product terms of all local peaks, and normalizing the sum by the main peak power to obtain the multimodality.

5. The method of claim 1, wherein, The calculation of the bypass risk index and the determination of the expected power comprise: based on the sparse I-V shape feature and the working condition data including the incident angle and the bypass threshold prediction, determining a sub-risk score, and weightedly synthesizing to generate the bypass risk index; when the bypass risk index exceeds the preset risk threshold, performing weighted summation on the local power peaks obtained from the sparse I-V shape feature to determine the expected power, wherein the weighted summation comprises: applying a softmax function to a weight input vector corresponding to each local power peak to generate a multi-peak weight, and weightedly summing the multiple local power peaks using the multi-peak weight; applying predefined weight coefficients to the step degree, the multimodality, the non-uniform severity, the incident angle and the local peak curvature to obtain the weight input vector.

6. The method of claim 1, wherein, It also includes determining the source of the non-uniform state, comprising: evaluating the frequency domain stability of the power sequence to obtain a frequency domain stability evaluation result, including: calculating the high frequency flicker index and the low frequency smoothing index of the power sequence; checking the monotonicity of the power loss with respect to the incident angle to obtain a monotonicity checking result, including: performing statistical coefficient test on the time series of the power loss and the incident angle; evaluating the consistency of the power loss and the reference coverage of the day to obtain a consistency evaluation result; According to the frequency domain stability evaluation result, the monotonicity test result and the consistency evaluation result, it is determined that the source of the non-uniform state is the shielding or the dust.

7. The method of claim 1, wherein, Fusing the expected power with the baseline power prediction based on the bypass risk index also includes a process of determining a fusion coefficient: Mapping the bypass risk index through a monotonically increasing function to obtain an initial fusion coefficient; When the consistency flag used to represent the stability of the expected power is false, limiting the initial fusion coefficient by a preset upper limit to generate a final fusion coefficient; The generation of the consistency flag includes: evaluating the fluctuation of the expected power in adjacent time windows, checking the stability of the effective peak list corresponding to a plurality of local power peaks in adjacent time windows, and determining the true or false of the consistency flag according to the fluctuation evaluation result and the stability check result.

8. The method of claim 1, wherein, Also includes: In the period when the bypass risk index is lower than the preset low risk threshold, calculate the prediction residual between the short-term prediction curve and the measured power; According to the prediction residual, the daily reference coverage used to represent the degree of pollution on the day is updated online, so that the updated daily reference coverage is used for calculation in the subsequent prediction period.

9. The method of claim 1, wherein, Also includes initializing the daily reference coverage in the starting window period of the photovoltaic system, including: Determine the initial estimated value according to the inverter scanning point set; Compensate the initial estimated value according to the rate of change of the component temperature to generate the daily reference coverage for subsequent steps.

10. The method of claim 1, wherein, Also includes applying physical constraints to the short-term prediction curve, and the physical constraints at least include one of the following: non-negative constraint, inverter rated power upper limit constraint and climbing rate upper limit constraint.

Images