Intelligent photovoltaic power station power generation capacity prediction method and system based on multi-source meteorological data fusion

Abstract

This invention discloses an intelligent prediction method and system for photovoltaic power generation based on multi-source meteorological data fusion, belonging to the field of photovoltaic power generation prediction technology. It collects meteorological observations, satellite cloud images, numerical weather predictions, historical power generation data of the power plant, and photovoltaic string operation data, and constructs a data foundation with unified time base alignment. Based on the power transfer relationship of partial shading of photovoltaic strings, it inverts the optical thickness profile of the cloud shadow edge and the cloud shadow motion vector along the shape of the string power decrease. It uses this as a short-term anchor point and performs cross-scale consistency fusion with the macroscopic cloud system movement trend to obtain a unified cloud dynamic state. It inputs a time-series prediction neural network to output hourly, daily, and weekly multi-time-scale power generation prediction results. It extrapolates along the motion vector at the array's geometric position to generate a power generation mutation warning containing mutation depth, and uses the deviation during the mutation period as feedback to correct the power transfer relationship. By treating the string array as a distributed optical observation network, it eliminates the need for additional meteorological sensors to predict the power generation mutation depth.

Description

[0001] This invention relates to the field of photovoltaic power generation prediction technology, specifically to a method and system for intelligent prediction of photovoltaic power generation based on the fusion of multi-source meteorological data. Background Technology

[0002] The power generation of photovoltaic (PV) power plants is affected by multiple factors, including sunlight intensity, temperature, and weather conditions. Accurate prediction of PV power generation is crucial for power plant operation scheduling and grid absorption. Traditional PV power generation prediction mainly relies on simple extrapolation methods based on historical averages, which struggle to handle drastic fluctuations in irradiance caused by rapid cloud movement, thus limiting prediction accuracy. To improve prediction accuracy, the industry has gradually introduced multi-source data fusion and deep learning methods.

[0003] Chinese patent application CN113128793A discloses a photovoltaic power prediction method and system based on multi-source data fusion. This method acquires historical power generation sequences and external meteorological data for the day to be predicted, inputting them into a convolutional neural network sub-prediction model, a long short-term memory network sub-prediction model, and an extreme gradient enhancement tree sub-prediction model for prediction. The weather is then categorized into sunny, cloudy, and overcast based on the cloud cover index of the day. The results of each sub-model are weighted and fused according to the weather type to obtain the final photovoltaic power prediction result. However, this scheme sums the photovoltaic power station power as a whole sequence and weights it with weather classification. The spatial output differences between photovoltaic strings during cloud movement are averaged out as interference in the whole-station summation process, thus failing to recover the spatial structure information of cloud cover. This means it can only predict based on satellite cloud images with coarse spatial resolution and temporal lag, and the whole-station sequence with lost spatial structure, making it difficult to predict the depth of sudden changes in power generation before they occur.

[0004] Chinese patent application CN101969207A discloses a photovoltaic ultra-short-term power prediction method combining satellite remote sensing and meteorological telemetry technologies. This method combines satellite remote sensing cloud images with real-time meteorological elements measured by ground-based automatic weather stations, utilizes cloud-tracking wind technology to track the movement trend of weather systems, and predicts solar irradiance and ambient temperature for the next 0 to 4 hours, inputting these predictions into a photovoltaic power generation model to output power. However, this scheme's cloud movement tracking remains at the satellite cloud image level, its spatial resolution is limited by satellite imaging accuracy and suffers from imaging and transmission time lags. Furthermore, it does not utilize the photovoltaic power station string array itself as a distributed optical observation source, failing to obtain the fine optical thickness distribution of cloud shadow edges at the ground array scale, thus limiting its ability to predict the depth of near-term power generation abrupt changes.

[0005] It is evident that existing photovoltaic (PV) power generation prediction technologies generally treat PV strings as passive power generation units, averaging the spatial power output differences between strings during cloud movement as noise or using it only for fault diagnosis. They fail to recognize that these spatial differences actually encode cloud shadow optical thickness information, which determines the depth of impending power generation abrupt changes. This neglected understanding results in existing solutions lacking the ability to predict the depth of power generation abrupt changes in advance, failing to provide accurate early warnings for power plant operation scheduling and grid absorption before these abrupt changes occur. How to extract cloud shadow depth information from the PV power plant's own string-level operational data without adding additional meteorological sensors, and thus achieve multi-timescale power generation prediction including abrupt change depth, has become a core problem urgently needing to be solved in this technical field. Summary of the Invention

[0006] To address the core technical bottleneck in existing photovoltaic power plant power generation prediction technologies, which use spatial power output differences between strings as noise elimination and thus fail to predict the depth of power generation abrupt changes before they occur, this invention provides an intelligent prediction method and system for photovoltaic power plant power generation based on multi-source meteorological data fusion. By treating the photovoltaic power plant string array itself as a distributed optical observation network, inverting the optical thickness profile of cloud shadow edges based on the power transfer relationship of partial string shading, and performing cross-scale consistent fusion with the macroscopic cloud system movement trends from satellite and numerical weather prediction, this invention achieves accurate prediction of photovoltaic power plant power generation, including the depth of abrupt changes, at multiple time scales, without adding additional meteorological sensors, based on the principle of cloud shadow optical thickness field inversion and cross-scale physical state sharing.

[0007] The technical solution of this invention is as follows:

[0008] The intelligent prediction method for photovoltaic power generation based on multi-source meteorological data fusion includes the following steps:

[0009] Step S1: Collect meteorological observation data, satellite cloud image data, numerical weather prediction data, power plant historical power generation data and photovoltaic string operation data, mark the array geometric position of each photovoltaic string, and align the above data with a unified time base to construct a power generation prediction data base.

[0010] Step S2: Based on the array geometric position of each photovoltaic string and the power falling edge shape in the photovoltaic string operation data, and according to the power transfer relationship of partial shading of the photovoltaic string, the optical thickness profile of the cloud shadow edge and the cloud shadow motion vector are obtained by inversion.

[0011] Step S3: The optical thickness profile at the edge of the cloud shadow and the cloud shadow motion vector are fused across scales with the macroscopic cloud system movement trend obtained from the satellite cloud image data and the numerical weather forecast data to obtain a unified cloud dynamic state.

[0012] Step S4: Input the unified cloud dynamics state, the meteorological observation data and the power station's historical power generation data into the time series prediction neural network, and output multi-time scale power generation prediction results, including hourly power generation prediction results, daily power generation prediction results and weekly power generation prediction results.

[0013] Step S5: Based on the extrapolation of the optical thickness profile of the cloud shadow edge and the cloud shadow motion vector at the geometric position of the array, a power generation mutation warning with mutation depth is generated before the power generation mutation arrives. The deviation of the mutation period between the historical power generation data of the power station and the power generation prediction results at multiple time scales is fed back to correct the power transfer relationship of the partial shading of the photovoltaic string.

[0014] This invention also provides a multi-source meteorological data fusion photovoltaic power plant power generation prediction system, comprising: a data base construction module, used to collect meteorological observation data, satellite cloud image data, numerical weather prediction data, historical power generation data of the power plant, and photovoltaic string operation data, mark the array geometric position of each photovoltaic string, and align the above data with a unified time base to construct a power generation prediction data base; a cloud shadow inversion module, used to invert the cloud shadow edge optical thickness profile and cloud shadow motion vector based on the array geometric position of each photovoltaic string and the power fall edge shape in the photovoltaic string operation data, and according to the partial shading power transfer relationship of the photovoltaic string; and a cross-scale fusion module, used to fuse the cloud shadow edge optical thickness profile and the cloud shadow motion vector with the data based on the satellite cloud image data, satellite cloud image data, numerical weather prediction data, historical power generation data of the power plant, and photovoltaic string operation data, to obtain the cloud shadow edge optical thickness profile and cloud shadow motion vector; and a cross-scale fusion module, used to fuse the cloud shadow edge optical thickness profile and cloud shadow motion vector with the data based on the satellite cloud image data, satellite cloud image data, numerical weather prediction data, historical power generation data of the power plant, and photovoltaic string operation data, to obtain the cloud shadow edge optical thickness profile and cloud shadow motion vector, and according to the partial shading power transfer relationship of the photovoltaic string. The cloud image data and the macroscopic cloud movement trend obtained from the numerical weather prediction data are fused across scales to obtain a unified cloud dynamic state. The multi-scale prediction module is used to input the unified cloud dynamic state, the meteorological observation data and the power station's historical power generation data into a time-series prediction neural network, and output multi-time-scale power generation prediction results. The mutation early warning and feedback module is used to generate a power generation mutation early warning with mutation depth before the power generation mutation arrives, based on the extrapolation of the cloud shadow edge optical thickness profile and the cloud shadow movement vector at the array's geometric position, and to feed back the mutation time deviation of the power station's historical power generation data relative to the multi-time-scale power generation prediction results to correct the partial shading power transfer relationship of the photovoltaic string.

[0015] The beneficial effects of this invention are as follows:

[0016] First, by inverting the shape of the power fall-off edge in the photovoltaic string operation data into the optical thickness profile of the cloud shadow edge and the cloud shadow motion vector based on the power transfer relationship of partial shading of the photovoltaic string, the spatial structure information of cloud shading is restored and utilized. The mechanism is as follows: the output power of the photovoltaic string under partial shading conditions has a nonlinear relationship with the shading ratio. The mapping of the power fall-off edge of a single string to the optical thickness profile of the cloud shadow is non-injective. However, strings distributed in different array geometric positions constitute a multi-view spatial observation of the same cloud shadow optical thickness field. By considering the set of power fall-off edge shapes of each string as a spatially sparse observation after the optical thickness field is transferred through the transfer relationship, the optical thickness profile of the cloud shadow edge can be uniquely inverted through spatially sparse deconvolution. Compared with the method of summing the power of the entire station and averaging out the spatial output difference, this scheme transforms the spatial difference between strings that was originally discarded into a usable cloud shadow depth measurement signal, and obtains a fine optical thickness distribution at the ground array scale without adding any additional meteorological sensing equipment.

[0017] Second, by using the inverted cloud shadow edge optical thickness profile and cloud shadow motion vector as short-term anchor points, and the macroscopic cloud system movement trend from satellite and numerical weather prediction as long-term drift, a unified cloud dynamic state shared by three time scales is obtained through cross-scale consistency fusion. The mechanism is as follows: the cloud dynamics obtained from string inversion have high spatial resolution but only cover the area above the power station and the immediate period; satellite cloud images and numerical weather predictions have a wide range and long duration but relatively coarse resolution and time lag. By constructing velocity field consistency constraints and optical thickness consistency constraints, the two converge to the same continuous cloud dynamic state. Short-term anchor points are used for local localization of long-term drift, and long-term drift extends short-term anchor points, thus establishing hourly, daily, and weekly predictions on the same physical cloud dynamics. Compared to the approach of only tracking cloud systems with cloud traces and winds at the satellite cloud image level, where each time scale is independent, this scheme's multi-scale predictions are no longer fragmented.

[0018] Third, the synergy between the cloud shadow edge optical thickness profile inversion feature and the cross-scale consistency fusion feature generates a synergistic effect exceeding the sum of their individual effects. Specifically: cloud shadow inversion alone cannot extrapolate high-resolution cloud dynamics to time periods beyond the hour; cross-scale fusion alone lacks fine optical thickness at the ground array scale as a calibration anchor point. With the synergy of the two, this scheme can extrapolate the optical thickness profile along the cloud shadow motion vector at the array's geometric position, predicting the depth of power generation mutations before they occur and generating early warnings of power generation mutations containing the mutation depth—a capability that neither feature alone can achieve. Furthermore, by feeding back the prediction deviation during mutation periods to correct the power transfer relationship of partial shading of photovoltaic strings, this transfer relationship continuously adapts to module aging and seasonal changes, forming a closed loop; and by separating the clear-sky output baseline offset into gradual and abrupt components, the inherent output differences caused by string dust accumulation, attenuation, and faults are removed before inversion, and string health, representing the impact of individual string anomalies on the entire station, is output simultaneously, allowing string anomaly identification to be naturally integrated into the main chain of power generation prediction rather than being an isolated diagnostic step. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method for predicting photovoltaic power generation by fusing multi-source meteorological data, as described in this invention.

[0020] Figure 2 This is an architecture diagram of the photovoltaic power generation prediction system based on multi-source meteorological data fusion, as described in this invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0022] This embodiment takes a centralized ground-mounted photovoltaic power station with an installed capacity of 100MW as an example. The power station includes several photovoltaic arrays, and each array consists of multiple photovoltaic strings that are combined through combiner boxes and then connected to an inverter. The entire station has a total of 1200 photovoltaic strings, and the rated power of each photovoltaic string is approximately 80kW. The method embodiment will be described in detail below according to steps S1 to S5, followed by the system embodiment.

[0023] Step S1: Collect multi-source data and construct a data foundation for power generation prediction. This step collects five types of data: meteorological observation data, satellite cloud image data, numerical weather prediction data, historical power generation data of the power plant, and photovoltaic string operation data. Specifically, meteorological observation data comes from the power plant's automatic weather station, including irradiance, ambient temperature, and wind speed and direction, with an irradiance sampling time resolution of 1 minute; satellite cloud image data comes from geostationary meteorological satellites, including cloud cover and cloud top height, with a time resolution of 15 minutes and a spatial resolution of approximately 1 km; numerical weather prediction data comes from the meteorological operational system, including irradiance forecasts for the future period, with a time resolution of 1 hour; historical power generation data of the power plant includes the historical output power of each photovoltaic string and the entire plant, with a time resolution of 1 minute; and photovoltaic string operation data includes the inverter output power of each photovoltaic string via the inverter and the combiner box status, with a time resolution of 1 minute.

[0024] After data collection, the array geometric position of each photovoltaic string is labeled. This array geometric position is represented by two-dimensional coordinates in the power station's planar coordinate system, denoted as the east and north coordinates of the string's junction point on the power station plane, in meters (m). The purpose of labeling the array geometric position is that when cloud shadows sweep across the power station from west to east, strings at different positions along the cloud shadow's leading edge will experience a sequential power decrease. The array geometric position provides a geometric reference for subsequently retrieving the spatial distribution of the cloud shadow from the power decrease of the strings.

[0025] Because the time resolutions of the five types of data mentioned above are inconsistent, this step unifies the time base of all source data, using 1 minute as the unified time resolution. Linear interpolation upsampling is applied to satellite cloud imagery data and numerical weather prediction data with coarser time resolutions, while nearest-neighbor resampling is used for data with finer time resolutions. All data is then aligned to a unified Coordinated Universal Time (UTC) timestamp, ensuring that each source data has a one-to-one corresponding value at the same time. The time-base-aligned source data is then organized into a power generation prediction data base based on photovoltaic string numbers and timestamps for use in subsequent steps.

[0026] To distinguish the power variations caused by cloud cover from the inherent power differences of the photovoltaic strings in subsequent steps, this step further establishes a clear-sky power output baseline for each photovoltaic string. The clear-sky power output baseline is the power that the string should output under clear-sky conditions without cloud cover, given the solar position and ambient temperature. It can be obtained by statistically analyzing the string's historical clear-sky operating data, categorized by solar altitude angle. Based on this, the photovoltaic string operating data is normalized using the clear-sky power output baseline of each photovoltaic string as a benchmark to obtain the normalized power, calculated as follows: ,in: For the first A photovoltaic string in The normalized power at time t is a scalar, and its value ranges from 0 to 1. , dimensionless, is calculated by this formula, representing the actual output ratio of the string relative to its own clear sky output level, thereby separating the inherent output differences of each photovoltaic string caused by orientation, tilt angle, component batch or aging degree. For the first A photovoltaic string in The inverter output power at any given time is a scalar quantity, and its value range is... The unit is kW, which is obtained from the photovoltaic string operation data collected in step S1; For the first A photovoltaic string in The baseline of clear-sky power output at any given time is a scalar, and its value range is... The unit is kW, and it is obtained by statistically analyzing the historical clear-sky operating data of this series of units according to the solar altitude angle. The rated power of this string is 80kW in this embodiment; subscript The value is the photovoltaic string number, and the range is: to integers, The total number of photovoltaic strings is 1200 in this embodiment. The normalized power decreases over time during the cloud shadow period, which is the shape of the normalized power fall-off edge reflecting the optical thickness at the edge of the cloud shadow.

[0027] Step S2: Invert the optical thickness profile at the cloud shadow edge and the cloud shadow motion vector. The clear-sky power output baseline itself shifts over time: string dust accumulation and component attenuation cause the baseline to slowly decrease, while string failures cause the baseline to suddenly drop. If baseline shift and cloud shadow occlusion are not distinguished, the inverted optical thickness will be contaminated by the baseline shift. Therefore, this step first separates the clear-sky power output baseline shift into gradual and abrupt components. Let the first component be... Each string in The clear-sky power output baseline offset at any given time is the difference between the observed clear-sky baseline and the nominal clear-sky baseline. The following separation problem is solved for this offset sequence:

[0028] ,

[0029] in: For the first Each string in The gradual component of the baseline offset of power output under clear skies at any given time is a scalar, and its value range is... The unit is kW, which is obtained by solving this minimization problem and represents the slow baseline drop caused by dust accumulation or component degradation. For the first Each string in The abrupt component of the baseline offset of power output under clear skies at any given time is a scalar, with a value range of [value missing]. The unit is kW, which is obtained by solving this minimization problem and represents the step baseline drop caused by string faults. For the first Each string in The baseline offset of the clear-sky power output at any given time is a scalar in kW, equal to the value of the string. The clear-sky baseline is obtained by subtracting the nominal clear-sky baseline from the observed clear-sky baseline, and is calculated from the clear-sky period data of that set in the data base. Here, is the smoothing regularization coefficient for the gradual component, and is a scalar with a range of values. The dimensionless multiplication with kW results in a dimensionless multiplication with kW. In this embodiment, it is set to 0.5. Its function is to constrain the second-order difference sparsity of the gradual component so as to maintain its smooth and gradual change. If the value is too large, the gradual component will be too flat and unable to track the true decay. If the value is too small, the gradual component will absorb abrupt components. Let be the sparsity regularization coefficient of the mutation component, and let be a scalar with a value range of . In this embodiment, 2.0 is used. Its function is to constrain the first-order difference sparsity of the mutation component so that a step occurs only at the fault time. If the value is too large, the fault will be missed; if the value is too small, noise will be misjudged as a fault. For absolute value operations; subscript To unify the time sequence number after time base alignment, the value range is: to integers, This represents the total number of samples within the statistical window. Based on the aforementioned gradual component, dust accumulation or module degradation in the photovoltaic (PV) strings is identified; based on the aforementioned abrupt component, PV string faults are identified. The gradual component is used to backfill and correct the clear-sky power output baseline, and the abrupt component is subtracted from the normalized power, ensuring that the inversion input only retains the power reduction caused by cloud shadows. Based on this, the string health score, representing the impact of individual PV string anomalies on the overall power generation of the power plant, is output. Its calculation is as follows:

[0030] ,

[0031] in: For the first The string health of each photovoltaic string is a scalar quantity with a value range of [value missing]. , dimensionless, is calculated by this formula, and characterizes the health of the string relative to its rated capacity. The closer it is to 1, the healthier it is. For the first The mean of the gradual components of each string within the statistical time window is a scalar in kW, which is obtained by averaging the gradual components obtained from solving the aforementioned separation problem over time. For the first The number of mutations identified in each string within the statistical time window is a scalar value, with a range of 1. to The integer, dimensionless, is obtained by counting the number of step points in the mutation component; Here, represents the mutation penalty weight, which is a scalar and has a value range of . , dimensionless, is taken as 0.5 in this embodiment to balance the contribution of gradual change and mutation to health; and The definition is the same as the formula mentioned above.

[0032] After baseline correction, this step performs the inversion of the optical thickness profile at the cloud shadow edge and the cloud shadow motion vector. It's important to note that a photovoltaic string consists of multiple photovoltaic modules connected in series. When a cloud shadow only partially blocks a portion of the modules in the string, the decrease in short-circuit current of the blocked modules forces the bypass diodes to conduct, causing the overall string output power to decrease nonlinearly with the proportion of blocking. This power transfer relationship is highly nonlinear. Therefore, the normalized power decrease of a single photovoltaic string at a given moment cannot uniquely determine the optical thickness profile of the cloud shadow blocking that string. In other words, the mapping of the normalized power decrease edge of a single string to the cloud shadow optical thickness profile is non-injective; multiple different optical thickness distributions may correspond to the same overall string power decrease. However, the 1200 photovoltaic strings distributed at different array geometric positions constitute multi-view, multi-moment spatial sampling of the same cloud shadow that moves over time: the edge of the same cloud shadow sweeps across multiple strings located on its path at different times, and each string records the power fall-off shape related to the local optical thickness of the cloud shadow edge when it is swept. By collecting these power fall-off shapes scattered in space and time into a set, the spatial constraints provided by the array geometry can be used to transform the ill-conditioned inverse problem for a single string into a well-posed inverse problem for the entire string array. Based on the above understanding, this step models the set of normalized power fall-off shapes of each photovoltaic string as a spatially sparse observation of the cloud shadow optical thickness field after the power transfer relationship of partial shading by the photovoltaic strings. Its forward observation model is as follows:

[0033] ,

[0034] in: For the first string The normalized power at time S1 is defined using the same formula as in step S1. The power transfer relationship of partial shading in a photovoltaic string is given by parameters. A defined monotonic nonlinear function, whose input is the effective transmittance of the string receiver and output is the normalized power, is dimensionless and mapped to dimensionless values. It is used to characterize the step-like power jump caused by the conduction of the bypass diode under partial obstruction. The parameters are... Feedback and corrections are provided in step S5; Let be a two-dimensional spatial location on the power station plane, and be a two-dimensional vector with units of m. Traversing the ... The light-receiving area of ​​each string ; For the first The array geometric position of each string is a two-dimensional vector in m, which is obtained by annotation in step S1; Let be the string space sampling kernel function, and let be a scalar with a range of values. , dimensionless, represents the light-receiving weight of spatial position relative to the center of the string, and its integral in the light-receiving region of the string is 1; for Time Power Station Planar Position The optical thickness of the cloud shadow at that location is a scalar quantity, with a value range of [value missing]. Dimensionless, to be solved by inversion. The corresponding atmospheric transmittance is given according to Beer-Lambert's law; For the first string The observation noise at any given time is a scalar with dimensionless units; subscript With symbols The definition is the same as the aforementioned formula. Since the cloud shadow edges are spatially concentrated and the cloud shadow optical thickness field is sparse along the spatial gradient, this step uses spatially sparse deconvolution to separate and inverse the cloud shadow optical thickness field from the set of shapes along the normalized power decrease. The inversion objective is:

[0035] ,

[0036] in: The cloud shadow optical thickness field is obtained by inversion, and is a matrix defined on the power station plane grid. Each element is dimensionless and is obtained by solving this minimization problem. Its value along the edge of the cloud shadow constitutes the cloud shadow edge optical thickness profile. Let be a spatial sampling operator, be a matrix, and be composed of the sampling kernel functions of each string group. The array is discretized and assembled, and the continuous optical thickness field is mapped to the effective transmittance observation at each string. For the first The effective transmittance at each string is a scalar, dimensionless. Let L2 be the norm of the vector; For spatial gradient operators, The L1 norm of the optical thickness field spatial gradient is used to constrain the spatial sparsity of cloud shadow edges. is the sparse regularization coefficient, and is a scalar with a range of values ​​of . The value is dimensionless. In this embodiment, it is set to 0.1. If the value is too large, the inverted cloud shadow edges will be too smooth and lose details. If the value is too small, the inversion results will be affected by noise and produce false edges. , , The definition is the same as the aforementioned formula. After obtaining the cloud shadow optical thickness field at different times through inversion, the cloud shadow motion vector is estimated by spatial registration of the optical thickness fields at adjacent times:

[0037] ,

[0038] in: Let be the cloud shadow motion vector, a two-dimensional vector, with a value range of . The unit is m / s, which is obtained by solving this minimization problem and represents the speed and direction of the cloud shadow's movement on the power station plane. for Time and location The optical thickness of the cloud shadow at that location is defined in the same way as the aforementioned inversion results; The time interval between adjacent optical thickness fields is denoted by , where is a scalar and its value ranges from . The unit is seconds (s), and in this embodiment, it is 60 seconds. and All are two-dimensional vectors with units of m. The above-mentioned inversion target and motion vector estimation are solved by an alternating iterative method: first, fix the motion vector and perform a sparse deconvolution update based on the near-end gradient on the optical thickness field; then fix the optical thickness field and perform a spatial registration update on the motion vector; this alternating iteration continues until the change in the optical thickness field between two adjacent iterations is lower than the preset convergence threshold. In this embodiment, the convergence threshold is 0.01, and the maximum number of iterations is 20. When multiple cloud shadows exist simultaneously above the power station, since the motion vectors and optical thickness profiles of each cloud shadow are different, the inverted optical thickness field presents multiple mutually separated high-value connected regions in space. This step divides the optical thickness field according to the connected regions, thereby separating and inverting the optical thickness profiles and motion vectors of the cloud shadow edges of each cloud shadow. The resulting cloud shadow edge optical thickness profiles finely depict the transition shape of the cloud shadow edge from thin to thick in space, while the cloud shadow motion vectors give the direction and velocity of the cloud shadow approaching each string of the power station. Together, they constitute high-resolution short-term anchor points for subsequent cross-scale fusion and abrupt change early warning.

[0039] Step S3: Cross-scale consistent fusion yields a unified cloud dynamic state. The cloud shadow edge optical thickness profile and cloud shadow motion vector obtained in step S2 have high spatial resolution at the ground array scale, but their effective coverage is limited to the area above the power station and the immediate period, serving as short-term anchor points. Satellite cloud imagery data and numerical weather prediction data cover a wide area and have long durations, but their spatial resolution is coarser and there is a time lag between imaging and computation. Based on satellite cloud imagery data and numerical weather prediction data, the macroscopic cloud system movement trend is obtained as long-term drift. This step performs cross-scale consistent fusion of the two, constructs velocity field consistency constraints and optical thickness consistency constraints, and solves for the unified cloud dynamic state:

[0040] ,

[0041] in: To unify the cloud dynamic state, it is a vector containing the position, velocity field, and optical thickness field parameters of the cloud system above the power station. It is obtained by solving this minimization problem and represents the continuous cloud dynamics shared by the three time scales. This is a time scale identifier, with a value of [value to be filled in]. , , , representing hourly, daily, and weekly timeframes, respectively; For the first The prior cloud dynamics state at each time scale is a vector, given by the numerical weather prediction data at that scale. For the first The weights of data items at each time scale are scalars, with values ​​ranging from 1 to 2. The value is dimensionless. In this embodiment, the values ​​for hourly, daily, and weekly scales are 0.5, 0.3, and 0.2, respectively, reflecting the higher weight of recent scales. For unified cloud dynamics state The derived velocity field is a two-dimensional vector with units of m / s; The cloud shadow motion vector inverted in step S2 serves as the anchor point for the velocity field consistency constraint, defined as before; Here, represents the weight of the velocity field consistency constraint, and is a scalar with a value range of . , dimensionless, is taken as 1.0 in this embodiment, and is used to make the velocity field of the unified cloud dynamic state converge toward the short-time anchor point; For unified cloud dynamics state The derived optical thickness field is a vector and is dimensionless. The cloud shadow edge optical thickness profile obtained from step S2 serves as the anchor point for optical thickness consistency constraints. Here, represents the weight for the optical thickness consistency constraint, and is a scalar with a value range of . The dimensionless value is taken as 0.8 in this embodiment. Through this consistent fusion, the short-term anchor point is localized to the long-term drift, and the long-term drift is extended to the short-term anchor point, so that the hourly, daily, and weekly time scales share the same continuous unified cloud dynamic state. The above minimization problem is a quadratic form with respect to the unified cloud dynamic state, which can be solved iteratively by the conjugate gradient method. During the solution, the prior state given by numerical weather prediction is initialized first, and then gradually converged to the unified cloud dynamic state compatible with the short-term anchor point under the guidance of velocity field consistency constraints and optical thickness consistency constraints. This unified cloud dynamic state physically corresponds to a cloud system that evolves continuously within the hour to week time span, avoiding the problems of independent modeling of each time scale and jumps and contradictions at the boundaries of adjacent scale predictions in the existing technology. When the short-term anchor point is temporarily unavailable due to night or missing satellite data, the cross-scale fusion automatically degenerates to be dominated by the long-term drift, thereby ensuring the continuity and stability of the unified cloud dynamic state when data conditions change.

[0042] Step S4: Multi-scale power generation prediction. This step inputs the unified cloud dynamics state obtained in Step S3, the meteorological observation data from Step S1, and the power plant's historical power generation data into a time-series prediction neural network, outputting multi-time-scale power generation prediction results. The time-series prediction neural network employs an attention-based time-series encoder, which concatenates the unified cloud dynamics state, meteorological observation data, and historical power plant data into a time-series input sequence using a unified time base. After extracting time-series dependency features through a multi-layer time-series encoder, it connects to three multi-time-scale output heads: hourly, daily, and weekly, outputting hourly, daily, and weekly power generation prediction results, respectively. In this embodiment, the time-series encoder is stacked in four layers, with eight attention heads per layer and a hidden dimension of 256. The hourly output head predicts power generation for the next hour, the daily output head predicts for the next day, and the weekly output head predicts for the next seven days. Since the three output heads share the same encoder and the same unified cloud dynamics state, the multi-time-scale prediction results are physically consistent. The time-series prediction neural network can also be a time-series encoder based on gated recurrent units, which also connects three multi-timescale output heads of hourly, daily and weekly levels, and is implemented in a similar way to the time-series encoder based on attention mechanism.

[0043] Step S5: Abrupt Change Warning and Feedback Correction. This step, based on the cloud shadow edge optical thickness profile and cloud shadow motion vector retrieved in Step S2, extrapolates at the array's geometric location to predict power generation abrupt changes and quantify the depth of the abrupt change. Specifically, the cloud shadow optical thickness field is extrapolated along the cloud shadow motion vector to its future values ​​at each string array's geometric location:

[0044] ,in: The extrapolated first The geometric position of each string array in the future The optical thickness of the cloud shadow at time t is a scalar, dimensionless quantity, and is calculated by this formula; The cloud shadow optical thickness field is the result of the inversion in step S2. The cloud shadow motion vector is defined as before; For the first The array geometric position of each string, in meters; Let be the extrapolated lead time, and be a scalar with a value range of . The unit is s. The time window for sudden change warning is set to 1800 seconds in this embodiment. The future power generation of the entire station is then predicted by extrapolating the optical thickness field, and the depth of power generation change is quantified using the following formula:

[0045] ,in: Let be the depth of the power generation mutation, and be a scalar with a value range of . , dimensionless, is calculated by this formula, and represents the maximum expected drop in the total power generation of the station relative to the rated capacity within the warning time window; for The actual power generation of the entire station at any given time is a scalar quantity in MW, obtained by summing the power of each string in the data base. Future obtained by extrapolation The predicted power generation of the entire station at any given time is a scalar quantity in MW, which is obtained by summing the predicted power corresponding to the extrapolated optical thickness field of each string. The rated capacity of the entire station is 100MW, which is a scalar quantity. The definition is the same as before. When the depth of power generation mutation exceeds the preset mutation depth threshold, a power generation mutation warning containing the mutation depth is generated before the power generation mutation arrives. In this embodiment, the preset mutation depth threshold is 0.2, that is, the warning is triggered when the expected drop exceeds 20% of the rated capacity. The generated power generation mutation warning includes three items: the expected arrival time of the mutation, the mutation depth, and the duration of the mutation calculated from the cloud shadow motion vector. The expected arrival time of the mutation is obtained by dividing the geometric distance from the edge of the cloud shadow along the motion vector to the geometric position of each string array by the velocity modulus of the cloud shadow motion vector, so that operators and dispatching departments can arrange the start-up of backup power and the allocation of load in advance. This step further pushes the prediction results of different time scales according to their timeliness: the hourly power generation prediction results and the power generation mutation warning are pushed to the power plant operators for real-time dispatching, and the daily power generation prediction results and weekly power generation prediction results are pushed to the grid dispatching department for maintenance planning and grid absorption planning, so that the power generation prediction of photovoltaic power plants is transformed from the traditional mode of relying on simple extrapolation to a high-resolution intelligent prediction mode based on multi-source meteorological data fusion.

[0046] To address extreme weather events such as sandstorms, snow accumulation, and smog, this step characterizes shading under these conditions as distributed continuous attenuation caused by aerosols or snow accumulation. Based on radiative transfer constraints and energy conservation constraints, a gray-box model is constructed to unify geometric shading formed by the optical thickness profile at the cloud shadow edge, distributed continuous attenuation, and photovoltaic module temperature attenuation into a physically constrained power generation field. Unlike geometric shading formed by the optical thickness at the cloud shadow edge, which exhibits a clear moving edge in space, the distributed continuous attenuation caused by sandstorms, snow accumulation, and smog is spatially widespread and gently distributed, changing slowly over time. These two types of attenuation manifest as different spatial frequency components in the optical thickness field. Therefore, this step assigns the high spatial frequency component that shifts with the cloud shadow motion vector in the inverted optical thickness field to the geometric shading optical thickness, and the slowly changing low spatial frequency background component to the distributed continuous attenuation optical thickness. The aerosol optical thickness and snow depth provided by numerical weather prediction data are used to calibrate this background component. The physically constrained power generation field is represented as follows:

[0047] ,

[0048] in: For a physically constrained power generation field in the first... string The output power at time t is a scalar quantity in kW, which is calculated using this formula. For the first The photoelectric conversion efficiency of each string is a scalar quantity, with a value ranging from 1 to 2. , dimensionless, determined by component parameters, and taken as 0.20 in this embodiment; For the first The light-receiving area of ​​each string is a scalar quantity, in units of... The size is determined by the component specifications and the string size; in this embodiment, it is taken as 400. ; for The irradiance reaching the module plane under unobstructed conditions at any given time is a scalar quantity, with units of . This was calculated using a model of solar position and clear-sky radiation. For the first The geometric occlusion optical thickness formed by the optical thickness profile of the cloud shadow edge at each group string is a scalar, dimensionless quantity, and is obtained by inversion in step S2. For the first The distributed, continuously decaying optical thickness at each string is formed by aerosols or snow cover. It is a scalar, dimensionless quantity, and is estimated from the aerosol optical thickness and snow depth data of numerical weather prediction. Let be the power temperature coefficient of the photovoltaic module, and let be a scalar with a value range of . The unit is The value is determined by the component parameters; in this embodiment, it is set to 0.004. For the first The temperature of each string component is a scalar quantity, in units of... It was estimated from meteorological observation data and component thermal models; The nominal operating temperature is the reference temperature; this is a scalar quantity, and the unit is... Take 25 The gray box model is constrained by the law of conservation of energy, meaning that the incident radiation reaching the component plane is equal to the sum of the transmitted radiation, absorbed radiation, and scattered radiation, thus constraining the range of values ​​for the geometric shading optical thickness and the distributed continuous attenuation optical thickness.

[0049] Finally, this step uses the deviation of historical power generation data from the multi-timescale power generation prediction results during abrupt changes to correct the power transfer relationship of partial shading of the photovoltaic strings. The abrupt change period is denoted as the set of times when the abrupt change is predicted, and the parameters of this transfer relationship are updated using the following formula:

[0050] ,

[0051] in: Power transfer relationship of partial shading of photovoltaic string The parameters are vectors, which are iteratively corrected by this update formula, representing the shape of the power step transition under partial shading; The learning rate is the feedback rate, which is a scalar and has a range of values. The value is dimensionless; in this embodiment, it is set to 0.01. If the value is too large, the parameter update will oscillate; if the value is too small, the convergence will be slow. This is a time set of mutation periods, consisting of the times when mutations are predicted in step S4. for The predicted power generation of the entire station at any given time is in MW, and the definition is the same as before; for The actual power generation of the entire station at any given time is a scalar quantity in MW, obtained from the historical power generation data of the power station. The gradient of the total station predicted power with respect to the transfer relationship parameters is obtained by inverse differentiation of the prediction model; This indicates parameter assignment and update. Through this feedback, the power transfer relationship of partial shading of photovoltaic strings continuously adapts to module aging and seasonal changes, ensuring consistency between cloud shadow inversion and abrupt change depth prediction throughout the entire life cycle of the power plant, forming a closed loop from power generation observation feedback to the transfer relationship.

[0052] The present invention also provides a photovoltaic power generation prediction system for multi-source meteorological data fusion, used to execute the above method embodiments, including a data base construction module, a cloud shadow inversion module, a cross-scale fusion module, a multi-scale prediction module, and a sudden change early warning and feedback module. Each module corresponds one-to-one with steps S1 to S5 of the method embodiments.

[0053] The data base construction module is used to collect meteorological observation data, satellite cloud image data, numerical weather prediction data, historical power generation data of the power plant, and photovoltaic string operation data. It marks the array geometric position of each photovoltaic string and aligns the above data with a unified time base to construct a power generation prediction data base. This module includes a data acquisition subunit, a time base alignment subunit, and a baseline establishment subunit. The data acquisition subunit collects the aforementioned five types of data via the power plant's automatic weather station, geostationary meteorological satellite receiving link, meteorological business system interface, power plant operation monitoring system, inverter, and combiner box. The time base alignment subunit resamples each source data to 1 minute and aligns the timestamps as described in step S1. The baseline establishment subunit establishes a clear-sky power output baseline for each photovoltaic string and performs normalization to obtain the normalized power falling edge shape. The implementation of this module is consistent with the content described in step S1 of the method embodiment.

[0054] The cloud shadow inversion module is used to invert the optical thickness profile of the cloud shadow edge and the cloud shadow motion vector based on the array geometry of each photovoltaic string and the power fall-edge shape in the photovoltaic string operation data, according to the power transfer relationship of partial shading of the photovoltaic strings. This module includes a baseline separation subunit, a health output subunit, and a sparse deconvolution subunit: the baseline separation subunit separates the clear-sky power output baseline offset into gradual and abrupt components as described in step S2 of the method embodiment, and adaptively corrects the clear-sky power output baseline accordingly; the health output subunit outputs the string health, representing the impact of individual photovoltaic string anomalies on the overall power generation of the power station; the sparse deconvolution subunit models the set of normalized power fall-edge shapes of each photovoltaic string as a spatially sparse observation of the cloud shadow optical thickness field after transfer relationships, and inverts the cloud shadow edge optical thickness profile and cloud shadow motion vector of multiple cloud shadows through spatially sparse deconvolution separation. The implementation of this module is consistent with the content described in step S2 of the method embodiment.

[0055] The cross-scale fusion module is used to perform cross-scale consistency fusion of the cloud shadow edge optical thickness profile and the cloud shadow motion vector with the macroscopic cloud system movement trend obtained based on the satellite cloud image data and the numerical weather prediction data, to obtain a unified cloud dynamic state. This module uses the cloud shadow edge optical thickness profile and the cloud shadow motion vector as short-term anchor points and the macroscopic cloud system movement trend as long-term drift, constructing velocity field consistency constraints and optical thickness consistency constraints as described in step S3 of the method embodiment, so that the hourly, daily, and weekly time scales share the same continuous unified cloud dynamic state. The implementation of this module is consistent with the content described in step S3 of the method embodiment.

[0056] The multi-scale prediction module is used to input the unified cloud dynamics state, the meteorological observation data, and the power station's historical power generation data into a time-series prediction neural network, and output multi-time-scale power generation prediction results, including hourly, daily, and weekly power generation prediction results. The time-series prediction neural network used in this module is either an attention-based time-series encoder or a gated recurrent unit-based time-series encoder, which connects to the hourly, daily, and weekly multi-time-scale output heads. The implementation of this module is consistent with step S4 of the method embodiment.

[0057] The mutation early warning and feedback module is used to generate a power generation mutation early warning with mutation depth based on the extrapolation of the optical thickness profile of the cloud shadow edge and the cloud shadow motion vector at the array's geometric position. This early warning generates a power generation mutation early warning before a mutation occurs and feeds back the deviation of the mutation period between the power plant's historical power generation data and the multi-timescale power generation prediction results to correct the power transfer relationship of the photovoltaic string partially blocked. This module includes an extrapolation early warning subunit, an extreme weather modeling subunit, and a feedback correction subunit: the extrapolation early warning subunit extrapolates the optical thickness field along the cloud shadow motion vector at the array's geometric position, quantifies the power generation mutation depth, and generates a power generation mutation early warning with mutation depth before the mutation occurs; the extreme weather modeling subunit constructs a gray box model based on radiative transfer constraints and energy conservation constraints under sandstorms, snow, or haze, unifying geometric shading, distributed continuous decay, and photovoltaic module temperature decay into a physically constrained power generation field; the feedback correction subunit feeds back the mutation period deviation to correct the power transfer relationship of the photovoltaic string partially blocked. Furthermore, this mutation early warning and feedback module also pushes hourly power generation forecast results to power plant operators for real-time dispatching, and pushes daily and weekly power generation forecast results to the grid dispatching department for maintenance planning and grid absorption planning. The implementation of this module is consistent with the content described in step S5 of the method embodiment.

[0058] Each module of the above system can be implemented by a computer program deployed on the power plant monitoring center server, or by a field-programmable logic device or an application-specific integrated circuit. The modules transmit power generation prediction data base, cloud shadow edge optical thickness profile, cloud shadow motion vector, unified cloud dynamic state and multi-time scale power generation prediction results through a data bus.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent prediction of photovoltaic power generation based on multi-source meteorological data fusion, characterized in that, include: Collect meteorological observation data, satellite cloud image data, numerical weather prediction data, historical power generation data of power plants and photovoltaic string operation data, mark the array geometric position of each photovoltaic string, and align the above data with a unified time base to construct a power generation prediction data base; Based on the array geometry of each photovoltaic string and the power fall-off shape in the photovoltaic string operation data, and according to the power transfer relationship of partial shading of the photovoltaic string, the optical thickness profile of the cloud shadow edge and the cloud shadow motion vector are obtained by inversion. The optical thickness profile at the edge of the cloud shadow and the cloud shadow motion vector are fused across scales with the macroscopic cloud system movement trend obtained from the satellite cloud image data and the numerical weather forecast data to obtain a unified cloud dynamic state. The unified cloud dynamics state, the meteorological observation data, and the power station's historical power generation data are input into a time-series prediction neural network, which outputs multi-time-scale power generation prediction results, including hourly power generation prediction results, daily power generation prediction results, and weekly power generation prediction results. Based on the extrapolation of the optical thickness profile of the cloud shadow edge and the cloud shadow motion vector at the geometric position of the array, a power generation mutation warning with mutation depth is generated before the power generation mutation arrives. The deviation of the mutation period between the historical power generation data of the power station and the power generation prediction results at the multi-time scale is fed back to correct the power transfer relationship of the partial shading of the photovoltaic string.

2. The method according to claim 1, characterized in that, The acquisition of the power fall-off shape includes: using the clear-sky power output baseline of each photovoltaic string as a reference, normalizing the photovoltaic string operation data, removing the inherent power output differences of each photovoltaic string, and obtaining the normalized power fall-off shape that reflects the optical thickness of the cloud shadow edge; the inherent power output differences are the power output differences caused by the orientation, tilt angle, module batch, or aging degree of the photovoltaic string.

3. The method according to claim 2, characterized in that, The offset of the clear-sky power output baseline is separated into a gradual component and a sudden component. The gradual component is used to identify dust accumulation or module degradation in the photovoltaic string, and the sudden component is used to identify faults in the photovoltaic string. Based on this, the clear-sky power output baseline is adaptively corrected, and the string health status, which characterizes the impact of individual photovoltaic string anomalies on the overall power generation of the power station, is output.

4. The method according to claim 3, characterized in that, The set of normalized power falling edge shapes of each photovoltaic string is modeled as a spatially sparse observation of the cloud shadow optical thickness field after being transmitted through the partial shading power transfer relationship of the photovoltaic string. The cloud shadow edge optical thickness profile and cloud shadow motion vector of each cloud shadow are separated and inverted from the set by spatially sparse deconvolution.

5. The method according to claim 4, characterized in that, The cross-scale consistency fusion includes: using the optical thickness profile of the cloud shadow edge and the cloud shadow motion vector as short-term anchor points and the macroscopic cloud system movement trend as long-term drift, constructing velocity field consistency constraints and optical thickness consistency constraints between scales, so that the hourly power generation prediction results, the daily power generation prediction results and the weekly power generation prediction results share the same continuous unified cloud dynamic state.

6. The method according to claim 5, characterized in that, In sandstorms, snow, or haze, shading is characterized as distributed continuous attenuation formed by aerosols or snow. A gray box model is constructed based on radiative transfer constraints and energy conservation constraints. The geometric shading formed by the optical thickness profile of the cloud shadow edge, the distributed continuous attenuation, and the temperature attenuation of the photovoltaic module are unified into a physically constrained power generation field. Based on this, the deviation between the partial shading power transfer relationship of the photovoltaic string and the relative nominal value of the temperature attenuation of the photovoltaic module is corrected.

7. The method according to claim 1, characterized in that, The meteorological observation data includes irradiance, ambient temperature, and wind speed and direction; the satellite cloud image data includes cloud cover and cloud top height; the numerical weather forecast data includes irradiance forecast; and the photovoltaic string operation data includes inverter output power and combiner box status. The unified time base alignment involves resampling the data from each source to the same time resolution and aligning the timestamps.

8. The method according to claim 1, characterized in that, The hourly power generation forecast results are pushed to the power plant operators for real-time dispatching, and the daily power generation forecast results and the weekly power generation forecast results are pushed to the power grid dispatching department for maintenance planning and power grid absorption planning.

9. The method according to claim 1, characterized in that, The time-series prediction neural network is a time-series encoder based on an attention mechanism or a time-series encoder based on a gated recurrent unit. The time-series encoder is connected to three multi-time-scale output heads: hourly, daily, and weekly.

10. A photovoltaic power plant power generation prediction system based on multi-source meteorological data fusion, used to implement the method described in any one of claims 1-9, characterized in that, include: The data base construction module is used to collect meteorological observation data, satellite cloud image data, numerical weather prediction data, power plant historical power generation data and photovoltaic string operation data, mark the array geometric position of each photovoltaic string, and align the above data with a unified time base to construct a power generation prediction data base. The cloud shadow inversion module is used to invert the optical thickness profile of the cloud shadow edge and the cloud shadow motion vector based on the array geometric position of each photovoltaic string and the power falling edge shape in the photovoltaic string operation data, according to the power transfer relationship of partial shading of the photovoltaic string. The cross-scale fusion module is used to perform cross-scale consistency fusion of the optical thickness profile of the cloud shadow edge and the cloud shadow motion vector with the macroscopic cloud system movement trend obtained based on the satellite cloud image data and the numerical weather forecast data to obtain a unified cloud dynamic state. The multi-scale prediction module is used to input the unified cloud dynamics state, the meteorological observation data and the power station's historical power generation data into the time-series prediction neural network, and output multi-time-scale power generation prediction results, including hourly power generation prediction results, daily power generation prediction results and weekly power generation prediction results. The mutation warning and feedback module is used to generate a power generation mutation warning with mutation depth before the power generation mutation arrives, based on the extrapolation of the optical thickness profile of the cloud shadow edge and the cloud shadow motion vector at the geometric position of the array, and to feed back the deviation of the mutation period of the historical power generation data of the power station relative to the power generation prediction results of the multi-time scale to correct the partial shading power transfer relationship of the photovoltaic string.

Images