Cleaning benefit prediction method and device based on load prediction and ash deposition analysis

Abstract

The embodiment of the invention provides a cleaning benefit prediction method and device based on load prediction and ash deposition analysis, and the method comprises the steps: obtaining the daily predicted meteorological data of the future N days from a meteorological prediction mechanism, and obtaining the daily inverter data of the latest historical N days from a station; according to the daily predicted meteorological data and the daily inverter data, on the basis of the first space-time diagram neural network model, predicting the daily predicted generating capacity of the future N days; according to the daily predicted meteorological data and the daily inverter data, based on the first double-layer LSTM neural network model, predicting a daily cleaning efficiency gain of the future N days; according to the daily predicted generating capacity, the daily cleaning efficiency gain and the cleaning cost of the future N days, predicting the predicted income of cleaning in the first day of the future N days; judging whether the predicted income of cleaning in the first day of the future N days is greater than the basic income or not; and if the earnings are greater than the basic earnings, cleaning is carried out in the first day of the next N days, otherwise, cleaning is not carried out.

Description

Technical Field

[0001] The invention relates to the field of photovoltaic power generation, and in particular to a cleaning benefit prediction method and device based on load prediction and dust accumulation analysis. Background Art

[0002] For photovoltaic power generation, due to environmental factors (such as wind, sand, dust, etc.), tiny particles are deposited on the surface of photovoltaic panels. These dusts may affect the light transmittance and photoelectric conversion efficiency of photovoltaic panels, thereby affecting the performance of photovoltaic power generation systems. At the same time, dusts will also cause the heat dissipation of photovoltaic panels to be blocked, and even produce hot spot effects. Some harsh environmental conditions will also cause chemical corrosion, which seriously affects the service life of photovoltaic panels. Therefore, cleaning photovoltaic panels can help improve power generation efficiency and extend service life. At present, the cleaning of photovoltaic panels mainly relies on checking the cleanliness of photovoltaic panels, and making regular operation and maintenance arrangements, and completing cleaning by manual or automated methods. The inventors have found that there are several major problems with photovoltaic cleaning at present, which often make the cleaning effect unsatisfactory. The cleaning cost is high. Whether it is manual cleaning or using automatic cleaning equipment for photovoltaic panels, the cost is relatively high. For a large number of small distributed photovoltaic power stations, they are often located in scattered locations, and operation and maintenance personnel have to spend a lot of money on the road. The complex on-site environment makes cleaning operations difficult. Blindly adopting this cleaning method may increase costs and affect economic benefits; the cleaning effect is greatly affected by the environment and the weather. If cleaning is carried out under continuous bad weather conditions, the benefits of cleaning may not cover the costs. At the same time, when weather conditions are not ideal, the cleaning effect may also be affected; the cleaning effect is difficult to quantify. The cleaning effect is affected by too many factors, and it is difficult to accurately quantify it, resulting in the inability to evaluate the pros and cons of the cleaning strategy, which is easy to cause unnecessary costs and economic losses.

[0003] In the process of implementing the present invention, the applicant discovered that there are at least the following problems in the prior art:

[0004] Traditional photovoltaic panel cleaning methods have the problems of high cleaning costs, great environmental impact, and difficulty in quantitative evaluation of the effects, and are unable to provide guidance for cleaning work arrangements. Summary of the invention

[0005] The embodiments of the present invention provide a cleaning benefit prediction method and device based on load prediction and dust accumulation analysis to solve the problems of traditional photovoltaic panel cleaning methods, such as high cleaning cost, great environmental impact of cleaning effects, difficulty in quantitative evaluation of cleaning effects, and inability to provide guidance for cleaning work arrangements.

[0006] To achieve the above objectives, on the one hand, an embodiment of the present invention provides a cleaning benefit prediction method based on load prediction and dust accumulation analysis, comprising:

[0007] Obtain daily forecast weather data for the next N days from weather forecast agencies, and obtain daily inverter data for the recent N days from stations;

[0008] According to the daily forecast meteorological data for the next N days and the daily inverter data for the recent N days, based on the first spatiotemporal graph neural network model, predict the daily forecast power generation for the next N days;

[0009] According to the daily forecast meteorological data for the next N days and the daily inverter data for the recent N days, based on the first double-layer LSTM neural network model, predict the daily cleaning efficiency gain for the next N days;

[0010] Based on the daily predicted power generation, daily cleaning efficiency gain, and cleaning cost for the next N days, the predicted benefits of cleaning on the first day of the next N days are predicted;

[0011] Determine whether the predicted return of cleaning on the first day of the next N days is greater than the basic return;

[0012] If the predicted return of cleaning on the first day of the next N days is greater than the basic return, cleaning will be performed on the first day of the next N days, otherwise no cleaning will be performed;

[0013] Among them, the daily cleaning efficiency gain represents the ratio of the daily predicted photovoltaic system efficiency after cleaning to the daily predicted photovoltaic system efficiency before cleaning; the daily predicted photovoltaic system efficiency after cleaning represents the daily predicted photovoltaic system efficiency for the next N days if cleaning is performed on the first day of the next N days; the daily predicted photovoltaic system efficiency before cleaning represents the daily predicted photovoltaic system efficiency for the next N days if no cleaning is performed in the next N days; the network parameters of the first spatiotemporal graph neural network model are pre-trained based on historical daily meteorological data, historical daily inverter data and historical daily power generation; the network parameters of the first two-layer LSTM neural network model are pre-trained based on historical daily meteorological data, historical daily inverter data, historical daily cleaning conditions and historical daily photovoltaic system efficiency.

[0014] Furthermore, the method of predicting the daily predicted power generation for the next N days based on the daily predicted meteorological data for the next N days and the daily inverter data for the recent N historical days based on the first spatiotemporal graph neural network model includes:

[0015] Encapsulating the daily forecast meteorological data for the next N days and the daily inverter data for the most recent N historical days into N first meteorological inverter data vectors, and sorting them in chronological order to obtain a first meteorological inverter data vector sequence;

[0016] Inputting the first meteorological inverter data vector sequence into a first spatiotemporal graph neural network model to predict daily power generation for the next N days;

[0017] Among them, the daily forecast meteorological data includes: weather type, cumulative radiation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and / or air quality index; daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency.

[0018] Further, the daily washing efficiency gain for the next N days is predicted based on the daily forecasted meteorological data for the next N days and the daily inverter data for the recent N days in history, based on the first double-layer LSTM neural network model, including:

[0019] Encapsulating the daily forecast meteorological data for the next N days, the daily cleaning mark, and the daily inverter data for the most recent N historical days into N second meteorological inverter data vectors, and sorting them in chronological order to obtain a second meteorological inverter data vector sequence;

[0020] The cleaning mark of the first second meteorological inverter data vector in the second meteorological inverter data vector sequence is set to be cleaned, and the cleaning marks of other second meteorological inverter data vectors are set to be not cleaned, and the second meteorological inverter data vector sequence with updated cleaning marks is input into the first double-layer LSTM neural network model to predict the daily photovoltaic system efficiency after cleaning for the next N days;

[0021] Setting the cleaning marks of all second meteorological inverter data vectors in the second meteorological inverter data vector sequence to not being cleaned, inputting the second meteorological inverter data vector sequence with the updated cleaning marks into the first double-layer LSTM neural network model, and predicting the un-cleaned daily photovoltaic system efficiency for the next N days;

[0022] Divide the daily PV system efficiency after cleaning by the daily PV system efficiency before cleaning for the next N days to obtain the daily cleaning efficiency gain for the next N days;

[0023] Among them, the daily forecast meteorological data includes: weather type, cumulative radiation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and / or air quality index; daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency.

[0024] Furthermore, the predicted benefit of cleaning on the first day of the next N days is predicted based on the daily predicted power generation, daily cleaning efficiency gain, and cleaning cost of the next N days, including:

[0025] The predicted post-cleaning power generation gain is calculated according to the following formula:

[0026] W=D1L1+…+D n L n (1)

[0027] Where W is the predicted power generation gain after cleaning, Dn is the daily cleaning efficiency gain on the nth day in the future, and L n is the daily predicted power generation for the nth day in the future;

[0028] The predicted return on cleaning on the first day of the next N days is calculated according to the following formula:

[0029] income=charge(W)-C (2)

[0030] Among them, income is the predicted income of cleaning on the first day of the next N days, charge() is the electricity price charging function provided by the plant to the owner, and C is the cleaning cost of one cleaning.

[0031] Furthermore, the method further comprises:

[0032] Obtain historical daily meteorological data, historical daily inverter data, and historical daily measured power generation for multiple days;

[0033] The historical daily meteorological data, the historical daily inverter data and the historical daily measured power generation of multiple days are used to form a first training vector on a daily basis, and multiple first training vectors corresponding to multiple days constitute a first training set; wherein the historical daily meteorological data and the historical daily inverter data are used as training data, and the historical daily measured power generation is used as label data;

[0034] Inputting the first training set into a second spatiotemporal graph neural network model, and predicting the historical daily predicted power generation of the corresponding day according to the historical daily meteorological data and the historical daily inverter data;

[0035] According to the historical daily predicted power generation and the historical daily measured power generation, a preset first loss function value is calculated, and according to the first loss function value, the model parameters of the second spatiotemporal graph neural network model are updated based on the gradient descent method until a preset first convergence condition is reached, and the model parameters of the second spatiotemporal graph neural network model are used to update the model parameters of the first spatiotemporal graph neural network model;

[0036] Wherein, a second spatiotemporal graph neural network model is pre-constructed, and the model structure of the second spatiotemporal graph neural network model is the same as that of the first spatiotemporal graph neural network model;

[0037] The historical daily meteorological data includes: weather type, radiation accumulation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and AQI; the historical daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency;

[0038] The format of the first training vector is:

[0039] Wherein, W is the weather type; R is the accumulated radiation, unit: megajoule / square meter; T is the temperature, unit: Celsius; H is the humidity, unit: %; D is the wind direction, unit: angle; V is the wind speed, meter / second; PM10 is the PM10 concentration, unit: %, PM2.5 is the PM2.5 content, unit: %; AQI is the air quality index, unit: microgram / cubic meter; are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter, G is the daily measured power generation, unit: kilowatt-hour.

[0040] Furthermore, the method further comprises:

[0041] Obtain multiple days of historical daily meteorological data, historical daily inverter data, historical daily cleaning data, and historical daily measured photovoltaic system efficiency;

[0042] The historical daily meteorological data, historical daily inverter data, historical daily cleaning and historical daily measured photovoltaic system efficiency of multiple days are used to form a second training vector on a daily basis, and multiple second training vectors corresponding to multiple days constitute a second training set; wherein the historical daily meteorological data, historical daily inverter data and historical daily cleaning are used as training data, and the historical daily measured photovoltaic system efficiency is used as label data;

[0043] The second training set is input into the second LSTM neural network model, and the photovoltaic system efficiency of the corresponding day is predicted based on the historical daily meteorological data, the historical daily inverter data and whether the historical daily cleaning data has been predicted;

[0044] According to the historical daily predicted photovoltaic system efficiency and the historical daily measured photovoltaic system efficiency, a preset second loss function value is calculated, and according to the second loss function value, the model parameters of the second LSTM neural network model are updated based on the gradient descent method until a preset second convergence condition is reached, and the model parameters of the second LSTM neural network model are used to update the model parameters of the first LSTM neural network model;

[0045] Wherein, a second LSTM neural network model is pre-constructed, and the model structure of the second LSTM neural network model is the same as that of the first LSTM neural network model;

[0046] The historical daily meteorological data includes: weather type, radiation accumulation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and AQI; the historical daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency;

[0047] The format of the second training vector is:

[0048] Wherein, W is the weather type; R is the accumulated radiation, unit: megajoule / square meter; T is the temperature, unit: Celsius; H is the humidity, unit: %; D is the wind direction, unit: angle; V is the wind speed, meter / second; PM10 is the PM10 concentration, unit: %, PM2.5 is the PM2.5 content, unit: %; AQI is the air quality index, unit: microgram / cubic meter; are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter, C is whether it has been cleaned on that day, and PR is the daily measured photovoltaic system efficiency.

[0049] On the other hand, an embodiment of the present invention provides a cleaning benefit prediction device based on load prediction and dust accumulation analysis, comprising:

[0050] A data acquisition unit is used to acquire daily forecasted meteorological data for the next N days from a meteorological forecasting agency, and to acquire daily inverter data for the most recent N days from a station;

[0051] A daily power generation prediction unit, configured to predict the daily predicted power generation for the next N days based on the daily predicted meteorological data for the next N days and the daily inverter data for the most recent N historical days, based on a first spatiotemporal graph neural network model;

[0052] A daily cleaning efficiency gain prediction unit, configured to predict the daily cleaning efficiency gain for the next N days based on the daily predicted meteorological data for the next N days and the daily inverter data for the recent N historical days, based on a first double-layer LSTM neural network model;

[0053] A revenue determination unit, for predicting a predicted revenue of cleaning on the first day of the next N days based on the predicted daily power generation, daily cleaning efficiency gain, and cleaning cost of the next N days;

[0054] The income comparison unit is used to determine whether the predicted income of cleaning on the first day of the next N days is greater than the basic income;

[0055] a cleaning scheme determination unit, configured to perform cleaning on the first day of the next N days if the predicted return of cleaning on the first day of the next N days is greater than the basic return, and not perform cleaning otherwise;

[0056] Among them, the daily cleaning efficiency gain represents the ratio of the daily predicted photovoltaic system efficiency after cleaning to the daily predicted photovoltaic system efficiency before cleaning; the daily predicted photovoltaic system efficiency after cleaning represents the daily predicted photovoltaic system efficiency for the next N days if cleaning is performed on the first day of the next N days; the daily predicted photovoltaic system efficiency before cleaning represents the daily predicted photovoltaic system efficiency for the next N days if no cleaning is performed in the next N days; the network parameters of the first spatiotemporal graph neural network model are pre-trained based on historical daily meteorological data, historical daily inverter data and historical daily power generation; the network parameters of the first two-layer LSTM neural network model are pre-trained based on historical daily meteorological data, historical daily inverter data, historical daily cleaning conditions and historical daily photovoltaic system efficiency.

[0057] Furthermore, the daily power generation prediction unit is configured as follows:

[0058] Encapsulating the daily forecast meteorological data for the next N days and the daily inverter data for the most recent N historical days into N first meteorological inverter data vectors, and sorting them in chronological order to obtain a first meteorological inverter data vector sequence; inputting the first meteorological inverter data vector sequence into a first spatiotemporal graph neural network model to predict the daily power generation for the next N days;

[0059] Among them, the daily forecast meteorological data includes: weather type, cumulative radiation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and / or air quality index; daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency.

[0060] Furthermore, the daily cleaning efficiency gain prediction unit comprises:

[0061] A data encapsulation module, used for encapsulating the daily forecast meteorological data for the next N days, the daily cleaning mark and the daily inverter data for the recent N historical days into N second meteorological inverter data vectors, and sorting them in chronological order to obtain a second meteorological inverter data vector sequence;

[0062] A post-cleaning daily photovoltaic system efficiency prediction module is used to set the cleaning mark of the first second meteorological inverter data vector in the second meteorological inverter data vector sequence to clean, and set the cleaning marks of other second meteorological inverter data vectors to not clean, and input the second meteorological inverter data vector sequence with the updated cleaning mark into the first double-layer LSTM neural network model to predict the post-cleaning daily photovoltaic system efficiency for the next N days;

[0063] An uncleaned daily photovoltaic system efficiency prediction module is used to set the cleaning marks of all second meteorological inverter data vectors in the second meteorological inverter data vector sequence to not be cleaned, input the second meteorological inverter data vector sequence with the updated cleaning mark to the first double-layer LSTM neural network model, and predict the uncleaned daily photovoltaic system efficiency for the next N days;

[0064] A daily cleaning efficiency gain determination module is used to divide the daily photovoltaic system efficiency after cleaning for the next N days by the daily photovoltaic system efficiency before cleaning to obtain the daily cleaning efficiency gain for the next N days;

[0065] Among them, the daily forecast meteorological data includes: weather type, cumulative radiation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and / or air quality index; daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency.

[0066] Furthermore, the revenue determination unit includes:

[0067] A post-cleaning power generation gain determination module, used to calculate the predicted post-cleaning power generation gain according to formula (1);

[0068] The profit determination module is used to calculate the predicted profit of cleaning on the first day of the next N days according to formula (2).

[0069] Furthermore, the device also includes: a second spatiotemporal graph neural network model training unit;

[0070] The second spatiotemporal graph neural network model training unit is configured as follows:

[0071] Obtain historical daily meteorological data, historical daily inverter data, and historical daily measured power generation for multiple days;

[0072] The historical daily meteorological data, the historical daily inverter data and the historical daily measured power generation of multiple days are used to form a first training vector on a daily basis, and multiple first training vectors corresponding to multiple days constitute a first training set; wherein the historical daily meteorological data and the historical daily inverter data are used as training data, and the historical daily measured power generation is used as label data;

[0073] Inputting the first training set into a second spatiotemporal graph neural network model, and predicting the historical daily predicted power generation of the corresponding day according to the historical daily meteorological data and the historical daily inverter data;

[0074] According to the historical daily predicted power generation and the historical daily measured power generation, a preset first loss function value is calculated, and according to the first loss function value, the model parameters of the second spatiotemporal graph neural network model are updated based on the gradient descent method until a preset first convergence condition is reached, and the model parameters of the second spatiotemporal graph neural network model are used to update the model parameters of the first spatiotemporal graph neural network model;

[0075] Wherein, a second spatiotemporal graph neural network model is pre-constructed, and the model structure of the second spatiotemporal graph neural network model is the same as that of the first spatiotemporal graph neural network model;

[0076] The historical daily meteorological data includes: weather type, radiation accumulation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and AQI; the historical daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency;

[0077] The format of the first training vector is:

[0078] Wherein, W is the weather type; R is the accumulated radiation, unit: megajoule / square meter; T is the temperature, unit: Celsius; H is the humidity, unit: %; D is the wind direction, unit: angle; V is the wind speed, meter / second; PM10 is the PM10 concentration, unit: %, PM2.5 is the PM2.5 content, unit: %; AQI is the air quality index, unit: microgram / cubic meter; are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter, G is the daily measured power generation, unit: kilowatt-hour.

[0079] Furthermore, the device also includes: a second LSTM neural network model training unit;

[0080] The second LSTM neural network model training unit is configured as follows:

[0081] Obtain multiple days of historical daily meteorological data, historical daily inverter data, historical daily cleaning data, and historical daily measured photovoltaic system efficiency;

[0082] The historical daily meteorological data, historical daily inverter data, historical daily cleaning and historical daily measured photovoltaic system efficiency of multiple days are used to form a second training vector on a daily basis, and multiple second training vectors corresponding to multiple days constitute a second training set; wherein the historical daily meteorological data, historical daily inverter data and historical daily cleaning are used as training data, and the historical daily measured photovoltaic system efficiency is used as label data;

[0083] The second training set is input into the second LSTM neural network model, and the photovoltaic system efficiency of the corresponding day is predicted based on the historical daily meteorological data, the historical daily inverter data and whether the historical daily cleaning data has been predicted;

[0084] According to the historical daily predicted photovoltaic system efficiency and the historical daily measured photovoltaic system efficiency, a preset second loss function value is calculated, and according to the second loss function value, the model parameters of the second LSTM neural network model are updated based on the gradient descent method until a preset second convergence condition is reached, and the model parameters of the second LSTM neural network model are used to update the model parameters of the first LSTM neural network model;

[0085] Wherein, a second LSTM neural network model is pre-constructed, and the model structure of the second LSTM neural network model is the same as that of the first LSTM neural network model;

[0086] The historical daily meteorological data includes: weather type, radiation accumulation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and AQI; the historical daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency;

[0087] The format of the second training vector is:

[0088] Wherein, W is the weather type; R is the accumulated radiation, unit: megajoule / square meter; T is the temperature, unit: Celsius; H is the humidity, unit: %; D is the wind direction, unit: angle; V is the wind speed, meter / second; PM10 is the PM10 concentration, unit: %, PM2.5 is the PM2.5 content, unit: %; AQI is the air quality index, unit: microgram / cubic meter; are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter, C is whether it has been cleaned on that day, and PR is the daily measured photovoltaic system efficiency.

[0089] The above technical solution has the following beneficial effects: predicting the future daily power generation according to the meteorological data and the inverter data through the first spatiotemporal graph neural network model, predicting the future daily cleaning efficiency gain according to the meteorological data and the inverter data through the first LSTM neural network model, judging whether the benefit of cleaning on the first day of the next N days is greater than the basic benefit according to the daily predicted power generation, daily cleaning efficiency gain, cleaning cost and basic benefit, and arranging the cleaning task accordingly. By evaluating whether to clean according to the power generation benefit and the cleaning cost, it not only comprehensively considers the weather factors represented by the meteorological data and the spatial factors of different sites represented by the inverter data of different sites, but also is simple, convenient and low-cost, effectively optimizes the judgment logic of when to clean the photovoltaic panels, provides the best cleaning solution, reduces unnecessary cleaning work, effectively avoids the additional costs and economic losses caused by poor cleaning effects due to seasons and weather, and maximizes the benefits of cleaning. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0091] Figure 1 It is a flow chart of a cleaning benefit prediction method based on load prediction and dust accumulation analysis according to one embodiment of the present invention;

[0092] Figure 2 This is a structural diagram of a cleaning benefit prediction device based on load prediction and dust accumulation analysis according to one embodiment of the present invention;

[0093] Figure 3 It is a schematic diagram of a graph neural network;

[0094] Figure 4 It is a schematic diagram of the spatiotemporal graph neural network;

[0095] Figure 5 It is a schematic diagram of the structure of a LTSM memory module;

[0096] Figure 6 It is a schematic diagram of the structure of a double-layer LSTM neural network consisting of two groups of LSTM neural networks;

[0097] Figure 7 This is a schematic diagram of the daily power generation data of the inverter at a certain station. DETAILED DESCRIPTION

[0098] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0099] On the one hand, if Figure 1 As shown, an embodiment of the present invention provides a cleaning benefit prediction method based on load prediction and dust accumulation analysis, comprising:

[0100] Step S10, obtaining daily forecasted meteorological data for the next N days from a meteorological forecasting agency, and obtaining daily inverter data for the most recent N days from a station;

[0101] Step S11, predicting the daily predicted power generation for the next N days based on the daily predicted meteorological data for the next N days and the daily inverter data for the most recent N historical days, based on the first spatiotemporal graph neural network model;

[0102] Step S12, predicting the daily washing efficiency gain for the next N days based on the daily forecast meteorological data for the next N days and the daily inverter data for the recent N historical days based on the first double-layer LSTM neural network model;

[0103] Step S13, predicting the predicted benefits of cleaning on the first day of the next N days based on the daily predicted power generation, daily cleaning efficiency gain, and cleaning cost of the next N days;

[0104] Step S14, determining whether the predicted return of cleaning on the first day of the next N days is greater than the basic return;

[0105] Step S15, if the predicted return of cleaning on the first day of the next N days is greater than the basic return, cleaning is performed on the first day of the next N days, otherwise cleaning is not performed;

[0106] Among them, the daily cleaning efficiency gain represents the ratio of the daily predicted photovoltaic system efficiency after cleaning to the daily predicted photovoltaic system efficiency before cleaning; the daily predicted photovoltaic system efficiency after cleaning represents the daily predicted photovoltaic system efficiency for the next N days if cleaning is performed on the first day of the next N days; the daily predicted photovoltaic system efficiency before cleaning represents the daily predicted photovoltaic system efficiency for the next N days if no cleaning is performed in the next N days; the network parameters of the first spatiotemporal graph neural network model are pre-trained based on historical daily meteorological data, historical daily inverter data and historical daily power generation; the network parameters of the first two-layer LSTM neural network model are pre-trained based on historical daily meteorological data, historical daily inverter data, historical daily cleaning conditions and historical daily photovoltaic system efficiency.

[0107] In some embodiments, the embodiments of the present invention are mainly based on load forecasting and dust accumulation analysis, introduce the calculation of cleaning costs, and optimize the cleaning plan. Load forecasting: generally refers to the prediction of the power generation of distributed photovoltaic power generation systems in the future. Since photovoltaics are most directly related to solar radiation intensity, a large amount of rich weather forecast data is involved. At present, the more mainstream load forecasting algorithms are mostly deep learning algorithms based on neural networks. Considering that distributed photovoltaic load forecasting may have highly complex spatiotemporal correlations, data from different sources have delays and other problems. The embodiment of the present invention adopts a spatiotemporal graph neural network model. Compared with the general neural network model, the spatiotemporal graph model can simultaneously derive time and space correlations, effectively reducing the problem of delays in data from different sources. Time correlation refers to the correlation of the observed values ​​of time series data in time, that is, the current value may be affected by past values. Spatial correlation refers to the correlation of data from adjacent geographical locations, and can also refer to the correlation between data from similar sources. In load forecasting, the spatiotemporal correlation between photovoltaic power generation and weather information and site location is relatively complex. For example, if the radiation and power generation of the previous hour are known, it is impossible to calculate the power generation of the next hour through the radiation of the next hour and a simple mathematical formula. It is also impossible to simply modify the existing prediction function of the inverter in a certain station for another inverter, although they are all intrinsically related. The delay of different source data refers to the data from the field collector and the data obtained from the weather forecast. There may be a delay in time and it is not possible to achieve a very accurate one-to-one correspondence at the same time. Especially for meteorological data, the cumulative radiation value of a certain hour may not be completely consistent with the actual situation. The complexity of spatiotemporal correlation and the delay of different source data have a certain negative impact on the prediction. The embodiment of the present invention uses the STGAT (Space-Time Graph Attention Network) model for load forecasting. Spatial-Temporal Graph Neutral Network (STGNN) is a graph neural network with time series. Each time step in STGNN is a graph, which is passed through the GCN / GAT network to obtain the resulting encoding graph that embeds the spatial interdependencies of the data. These encoding graphs can be modeled like time series data, and only the integrity of the graph structure of the data at each time step needs to be preserved. t ∈R N×F |t=0,…,T}, where N is the number of spatial nodes and F is the feature dimension. For such data, the space-time graph is an effective structure that can characterize the relationship between different nodes within a certain space-time range. The space-time graph can be expressed as: G t =(V,Et ,A t ), where V is the node set, E and A are the edge set and adjacency matrix. In general, the weather forecast data needs to support at least 14 days in the future, denoted as n, so that the forecast load data corresponding to n days can be generated respectively, that is, L = (L1, L2, ..., L n ), L n Represents the predicted load data for the nth day. The degree of PV dust accumulation varies significantly depending on the local air environment and climate conditions. In the relatively humid environment of southern China, the dust loss over time is not obvious compared to the arid environment. At the same time, after cleaning, the low dust loss is maintained for a longer time. In the arid environment of northern China, even after a long period of time, the dust on the PV panel will continue to accumulate significantly. Even if the dust has a considerable thickness, it will form a mechanical structure that tends to be stable, rather than becoming thinner or being blown away by the wind as expected. Cleaning or precipitation is the best way to reduce PV dust loss. There is no accurate definition of the degree of PV dust accumulation. In general, the PR value is used as a description of PV power generation efficiency. The PR value is the efficiency of the PV system (Performance Ratio), a key indicator for evaluating the quality of a PV system. It is the ratio of the actual output power of the power station to the theoretical output power, reflecting the proportional relationship between the actual input power of the entire power station to the grid after deducting all losses (including irradiation loss, line loss, device loss, dust loss, heat loss, etc.). Its standardized calculation formula (5) is generally carried out in accordance with the provisions of IEC61724. As the dust accumulation increases, the photovoltaic power generation efficiency gradually decays, and its PR value will continue to decline. For such complex nonlinear sequence problems with a long time, the LTSM long short-term memory neural network is the first choice for prediction. Compared with the RNN long short-term memory network, the LSTM (Long Short Term Memory Network) long short-term memory neural network has the advantages of better information storage and more accurate access to historical information, and can avoid the gradient disappearance problem in the RNN network. After the PR value of the cleaning and non-cleaning conditions is predicted by the LTSM neural network, the daily cleaning efficiency gain can be predicted based on the PR value of the cleaning and non-cleaning conditions; according to the daily predicted power generation, daily cleaning efficiency gain, cleaning cost and basic income for the next N days, it can be judged whether the income of cleaning on the first day of the next N days is greater than the basic income. The cleaning cost and basic income are both predictable. In this way, it is possible to determine the possible income of cleaning, and whether it can exceed the basic income. Only when it exceeds the basic income is it suitable to arrange personnel for cleaning.

[0108] The embodiment of the present invention has the following technical effects: predicting the future daily power generation according to meteorological data through the first spatiotemporal graph neural network model, predicting the future daily cleaning efficiency gain according to meteorological data through the first LSTM neural network model, judging whether the benefit of cleaning on the first day of the next N days is greater than the basic benefit according to the daily predicted power generation, daily cleaning efficiency gain, cleaning cost and basic benefit, and arranging the cleaning task accordingly. By evaluating whether to clean according to the power generation benefit and cleaning cost, it is simple, convenient and low-cost, effectively optimizing the judgment logic of when to clean the photovoltaic panels, providing the best cleaning plan, reducing unnecessary cleaning work, effectively avoiding the additional costs and economic losses caused by poor cleaning effects due to seasons and weather, and maximizing the benefits brought by cleaning.

[0109] Furthermore, the method of predicting the daily predicted power generation for the next N days based on the daily predicted meteorological data for the next N days and the daily inverter data for the recent N historical days based on the first spatiotemporal graph neural network model includes:

[0110] Encapsulating the daily forecast meteorological data for the next N days and the daily inverter data for the most recent N historical days into N first meteorological inverter data vectors, and sorting them in chronological order to obtain a first meteorological inverter data vector sequence;

[0111] Inputting the first meteorological inverter data vector sequence into a first spatiotemporal graph neural network model to predict daily power generation for the next N days;

[0112] Among them, the daily forecast meteorological data includes: weather type, cumulative radiation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and / or air quality index; daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency.

[0113] In some embodiments, the future daily forecast meteorological data generally needs to be obtained directly from meteorological forecast related agencies. It mainly includes daily weather types (including but not limited to sunny, cloudy, cloudy and rainy, etc.), radiation accumulation (unit: MJ / m2), temperature (unit: ℃), humidity (unit %), wind direction (unit: angle), wind speed (unit: m / s), PM10 concentration (unit: %), PM2.5 content (unit: %), AQI index (single air quality index, AirQuality Index, unit: μg / m3). The inverter data still uses the daily inverter data of the last N days in history. The inverter money of different sites reflects the difference in spatial location. Organize the daily forecast meteorological data and daily inverter data into the first meteorological inverter data vector form in chronological order: Among them, W is the weather type, such as sunny, cloudy, overcast, etc. When used as input data, it needs to be converted into a floating point number, such as sunny = 1, cloudy = 2, etc. R, T, H, D, V, PM10, PM2.5, AQI correspond to the cumulative radiation (MJ / m2), temperature (℃), humidity (%), wind direction (angle), wind speed (m / s), PM10 concentration (%), PM2.5 content (%), AQI (Air Quality Index, Air Quality Index, μg / m3) indicators, respectively. are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter. N first meteorological inverter data vectors can be obtained, and a first meteorological inverter data vector sequence is formed in chronological order. The first spatio-temporal graph neural network model is a spatio-temporal graph neural network (STGNN), and the network parameters in the first spatio-temporal graph neural network model are obtained by pre-training.

[0114] The embodiment of the present invention has the following technical effects: Distributed photovoltaic load forecasting may have highly complex spatiotemporal correlations, data from different sources may have delays, etc. The embodiment of the present invention adopts a spatiotemporal graph neural network model. Compared with a general neural network model, the spatiotemporal graph model can simultaneously derive temporal and spatial correlations, effectively reducing the problem of data delays from different sources.

[0115] Further, the daily washing efficiency gain for the next N days is predicted based on the daily forecasted meteorological data for the next N days and the daily inverter data for the recent N days in history, based on the first double-layer LSTM neural network model, including:

[0116] Encapsulating the daily forecast meteorological data for the next N days, the daily cleaning mark, and the daily inverter data for the most recent N historical days into N second meteorological inverter data vectors, and sorting them in chronological order to obtain a second meteorological inverter data vector sequence;

[0117] The cleaning mark of the first second meteorological inverter data vector in the second meteorological inverter data vector sequence is set to be cleaned, and the cleaning marks of other second meteorological inverter data vectors are set to be not cleaned, and the second meteorological inverter data vector sequence with updated cleaning marks is input into the first double-layer LSTM neural network model to predict the daily photovoltaic system efficiency after cleaning for the next N days;

[0118] Setting the cleaning marks of all second meteorological inverter data vectors in the second meteorological inverter data vector sequence to not being cleaned, inputting the second meteorological inverter data vector sequence with the updated cleaning marks into the first double-layer LSTM neural network model, and predicting the un-cleaned daily photovoltaic system efficiency for the next N days;

[0119] Divide the daily PV system efficiency after cleaning by the daily PV system efficiency before cleaning for the next N days to obtain the daily cleaning efficiency gain for the next N days;

[0120] Among them, the daily forecast meteorological data includes: weather type, cumulative radiation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and / or air quality index; daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency.

[0121] In some embodiments, the future daily forecast meteorological data generally needs to be obtained directly from meteorological forecast related agencies. It mainly includes daily weather types (including but not limited to sunny, cloudy, cloudy and rainy, etc.), radiation accumulation (unit: MJ / m2), temperature (unit: ℃), humidity (unit %), wind direction (unit: angle), wind speed (unit: m / s), PM10 concentration (unit: %), PM2.5 content (unit: %), AQI index (single air quality index, AirQuality Index, unit: μg / m3). The inverter data still uses the daily inverter data of the last N days. The inverter money of different sites reflects the difference in spatial location. The daily forecast meteorological data, daily cleaning mark and daily inverter data are organized into a second meteorological inverter data vector form in chronological order:

[0122] Among them, W is the weather type, such as sunny, cloudy, overcast, etc. When used as input data, it needs to be converted into a floating point number, such as sunny = 1, cloudy = 2, etc. R, T, H, D, V, PM10, PM2.5, AQI correspond to the cumulative radiation (MJ / m2), temperature (℃), humidity (%), wind direction (angle), wind speed (m / s), PM10 concentration (%), PM2.5 content (%), AQI (Air Quality Index, Air Quality Index, μg / m3) indicators, respectively. are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter, C indicates whether the corresponding day was cleaned, 1 can be used to indicate cleaning, and 0 to indicate no cleaning. N second meteorological inverter data vectors can be obtained, corresponding to the next N days, and forming a second meteorological inverter data vector sequence in chronological order. The first double-layer LSTM neural network model consists of two LSTM (Long Short-Term Memory) neural network models as follows Figure 6 The network parameters in the first double-layer LSTM neural network model are obtained by pre-training. The daily cleaning mark in the first day of the next N days (i.e., the first second meteorological inverter data vector in the second meteorological inverter data vector sequence) is set to after cleaning, and the second meteorological inverter data vector sequence is input to the first double-layer LSTM neural network model to predict the daily photovoltaic system efficiency after cleaning for the next N days; the daily cleaning marks of all N days (all the second meteorological inverter data vectors in the second meteorological inverter data vector sequence) are set to not cleaning, and the second meteorological inverter data vector sequence is input to the first double-layer LSTM neural network model again to predict the daily photovoltaic system efficiency without cleaning for the next N days. The daily cleaning efficiency gain is the daily photovoltaic system efficiency after cleaning of the corresponding day divided by the daily photovoltaic system efficiency without cleaning. The daily cleaning efficiency gain reflects the comparison of the efficiency of the photovoltaic system with cleaning relative to that without cleaning, and is used as a quantitative result of dust accumulation analysis; in some embodiments, the degree of photovoltaic dust accumulation varies significantly according to the local air environment and climatic conditions. In the relatively humid environment of southern China, the dust loss over time is not obvious compared to the arid environment. At the same time, after cleaning, the low dust loss is maintained for a longer time. In the arid environment such as northern China, even after a long period of time, the dust on the photovoltaic panel will continue to accumulate significantly. Even if the dust has a considerable thickness, it will form a mechanical mechanism that tends to a steady state, rather than becoming thinner or being blown away by the wind as hoped. Cleaning or precipitation is the best way to reduce photovoltaic dust loss. There is no accurate definition of the degree of photovoltaic dust accumulation, and the PR value can be used as a description of photovoltaic power generation efficiency. As the degree of dust accumulation deepens, the photovoltaic power generation efficiency gradually decays, and its PR value will continue to decline. For such complex and long-term nonlinear sequence problems, the LTSM long short-term memory neural network is preferred for prediction. Compared with RNN, LSTM (Long Short Term Memory Network) has the advantages of better information storage and more accurate access to historical information, and can avoid the gradient vanishing problem in RNN network. The embodiment of the present invention also defines a daily cleaning efficiency gain, which can reflect the comparison of the efficiency of the photovoltaic system with cleaning and without cleaning.

[0123] Furthermore, the predicted benefit of cleaning on the first day of the next N days is predicted based on the daily predicted power generation, daily cleaning efficiency gain, and cleaning cost of the next N days, including:

[0124] Calculate the predicted post-cleaning power generation gain W according to formula (1);

[0125] According to formula (2), calculate the predicted income of cleaning on the first day of the next N days;

[0126] In some embodiments, the daily predicted power generation (predicted load) of the same day is multiplied by the daily cleaning efficiency gain, and the products obtained for all N days are added together to obtain the predicted power generation gain after cleaning. The charge function is used to convert the power generation gain after cleaning into power generation revenue. The power generation revenue is subtracted from the cleaning cost of the first day of the next N days. The difference obtained is the predicted revenue of cleaning on the first day of the next N days. Among them, the cleaning cost of each day is predictable, and can be specifically predetermined based on the location of the station, labor costs, cleaning tool and material consumption, etc.; the charge function represents the electricity price charging standard provided by the station to the owner, which may be different for different projects in different regions and is predetermined between the station and the owner. The predicted income is determined based on the daily predicted power generation, daily cleaning efficiency gain, and cleaning cost for the next N days. Combined with the preset basic income, it can be determined whether the predicted income for cleaning on the first day of the next N days is greater than the basic income, that is, the difference f(W) between the predicted income and the basic income for cleaning on the first day of the next N days is obtained by subtracting the basic income from the predicted income income. If the value of f(W) is greater than 0, the income for cleaning on the first day of the next N days is greater than the basic income, otherwise the income for cleaning on the first day of the next N days is not greater than the basic income. If the predicted income is greater than the basic income, then the technicians can be arranged to clean on the first day of the next N days, otherwise it is not worth arranging the cleaning.

[0127] The embodiments of the present invention have the following technical effects: by calculating the cleaning revenue, it can be determined whether the cleaning revenue is greater than the basic revenue, and the cleaning plan is determined by comparing the cleaning revenue with the basic revenue, so that the selected cleaning time can more effectively improve the overall revenue compared to other times, reduce the success increase brought by other cleaning plans in the next N days, and collect the following problems.

[0128] Furthermore, the method further comprises:

[0129] Obtain historical daily meteorological data, historical daily inverter data, and historical daily measured power generation for multiple days;

[0130] The historical daily meteorological data, the historical daily inverter data and the historical daily measured power generation of multiple days are used to form a first training vector on a daily basis, and multiple first training vectors corresponding to multiple days constitute a first training set; wherein the historical daily meteorological data and the historical daily inverter data are used as training data, and the historical daily measured power generation is used as label data;

[0131] Inputting the first training set into a second spatiotemporal graph neural network model, and predicting the historical daily predicted power generation of the corresponding day according to the historical daily meteorological data and the historical daily inverter data;

[0132] According to the historical daily predicted power generation and the historical daily measured power generation, a preset first loss function value is calculated, and according to the first loss function value, the model parameters of the second spatiotemporal graph neural network model are updated based on the gradient descent method until a preset first convergence condition is reached, and the model parameters of the second spatiotemporal graph neural network model are used to update the model parameters of the first spatiotemporal graph neural network model;

[0133] Wherein, a second spatiotemporal graph neural network model is pre-constructed, and the model structure of the second spatiotemporal graph neural network model is the same as that of the first spatiotemporal graph neural network model;

[0134] The historical daily meteorological data includes: weather type, radiation accumulation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and AQI; the historical daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency;

[0135] The format of the first training vector is:

[0136] Wherein, W is the weather type; R is the accumulated radiation, unit: megajoule / square meter; T is the temperature, unit: Celsius; H is the humidity, unit: %; D is the wind direction, unit: angle; V is the wind speed, meter / second; PM10 is the PM10 concentration, unit: %, PM2.5 is the PM2.5 content, unit: %; AQI is the air quality index, unit: microgram / cubic meter; are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter, G is the daily measured power generation, unit: kilowatt-hour.

[0137] In some embodiments, a second spatiotemporal graph neural network model can be pre-constructed, and historical data such as historical daily meteorological data, historical daily inverter data, and historical daily measured power generation can be collected. The model parameters of the second spatiotemporal graph neural network model are trained and updated using the collected historical data. After the second spatiotemporal graph neural network model converges, the model parameters of the second spatiotemporal graph neural network model are used to update the model parameters of the first spatiotemporal graph neural network model.

[0138] Furthermore, the method further comprises:

[0139] Obtain multiple days of historical daily meteorological data, historical daily inverter data, historical daily cleaning data, and historical daily measured photovoltaic system efficiency;

[0140] The historical daily meteorological data, historical daily inverter data, historical daily cleaning and historical daily measured photovoltaic system efficiency of multiple days are used to form a second training vector on a daily basis, and multiple second training vectors corresponding to multiple days constitute a second training set; wherein the historical daily meteorological data, historical daily inverter data and historical daily cleaning are used as training data, and the historical daily measured photovoltaic system efficiency is used as label data;

[0141] The second training set is input into the second LSTM neural network model, and the photovoltaic system efficiency of the corresponding day is predicted based on the historical daily meteorological data, the historical daily inverter data and whether the historical daily cleaning data has been predicted;

[0142] According to the historical daily predicted photovoltaic system efficiency and the historical daily measured photovoltaic system efficiency, a preset second loss function value is calculated, and according to the second loss function value, the model parameters of the second LSTM neural network model are updated based on the gradient descent method until a preset second convergence condition is reached, and the model parameters of the second LSTM neural network model are used to update the model parameters of the first LSTM neural network model;

[0143] Among them, a second LSTM neural network model is pre-built, and the model structure of the second LSTM neural network model is the same as that of the first LSTM neural network model; the historical daily meteorological data includes: weather type, radiation accumulation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and AQI; the historical daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency; the format of the second training vector is:

[0144] Wherein, W is the weather type; R is the accumulated radiation, unit: megajoule / square meter; T is the temperature, unit: Celsius; H is the humidity, unit: %; D is the wind direction, unit: angle; V is the wind speed, meter / second; PM10 is the PM10 concentration, unit: %, PM2.5 is the PM2.5 content, unit: %; AQI is the air quality index, unit: microgram / cubic meter; are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter, C is whether it has been cleaned on that day, and PR is the daily measured photovoltaic system efficiency.

[0145] In some embodiments, a second LSTM neural network model can be pre-constructed, and historical data such as historical daily meteorological data, historical daily inverter data, historical daily cleaning data, and historical daily measured photovoltaic system efficiency can be collected. The model parameters of the second LSTM neural network model are trained and updated using the collected historical data. After the second LSTM neural network model converges, the model parameters of the second LSTM neural network model are used to update the model parameters of the first LSTM neural network model.

[0146] On the other hand, Figure 2 As shown, an embodiment of the present invention provides a cleaning benefit prediction device based on load prediction and dust accumulation analysis, comprising:

[0147] The data acquisition unit 200 is used to acquire the daily forecasted meteorological data for the next N days from the meteorological forecast agency, and to acquire the daily inverter data for the recent N days from the station;

[0148] A daily power generation prediction unit 201 is used to predict the daily predicted power generation for the next N days based on the daily predicted meteorological data for the next N days and the daily inverter data for the recent N historical days, based on a first spatiotemporal graph neural network model;

[0149] A daily cleaning efficiency gain prediction unit 202 is used to predict the daily cleaning efficiency gain for the next N days based on the daily predicted meteorological data for the next N days and the daily inverter data for the recent N historical days, based on a first double-layer LSTM neural network model;

[0150] The benefit determination unit 203 is used to predict the predicted benefit of cleaning on the first day of the next N days based on the daily predicted power generation, daily cleaning efficiency gain, and cleaning cost of the next N days;

[0151] The income comparison unit 204 is used to determine whether the predicted income of cleaning on the first day of the next N days is greater than the basic income;

[0152] A cleaning scheme determination unit 205 is configured to perform cleaning on the first day of the next N days if the predicted profit of cleaning on the first day of the next N days is greater than the basic profit, and not perform cleaning otherwise;

[0153] Among them, the daily cleaning efficiency gain represents the ratio of the daily predicted photovoltaic system efficiency after cleaning to the daily predicted photovoltaic system efficiency before cleaning; the daily predicted photovoltaic system efficiency after cleaning represents the daily predicted photovoltaic system efficiency for the next N days if cleaning is performed on the first day of the next N days; the daily predicted photovoltaic system efficiency before cleaning represents the daily predicted photovoltaic system efficiency for the next N days if no cleaning is performed in the next N days; the network parameters of the first spatiotemporal graph neural network model are pre-trained based on historical daily meteorological data, historical daily inverter data and historical daily power generation; the network parameters of the first double-layer LSTM neural network model are pre-trained based on historical daily meteorological data, historical daily inverter data, historical daily cleaning conditions and historical daily photovoltaic system efficiency.

[0154] Furthermore, the daily power generation prediction unit 201 is configured as follows:

[0155] Encapsulating the daily forecast meteorological data for the next N days and the daily inverter data for the most recent N historical days into N first meteorological inverter data vectors, and sorting them in chronological order to obtain a first meteorological inverter data vector sequence;

[0156] Inputting the first meteorological inverter data vector sequence into a first spatiotemporal graph neural network model to predict daily power generation for the next N days;

[0157] Among them, the daily forecast meteorological data includes: weather type, cumulative radiation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and / or air quality index; daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency.

[0158] Furthermore, the daily cleaning efficiency gain prediction unit 202 includes:

[0159] A data encapsulation module, used for encapsulating the daily forecast meteorological data for the next N days, the daily cleaning mark and the daily inverter data for the recent N historical days into N second meteorological inverter data vectors, and sorting them in chronological order to obtain a second meteorological inverter data vector sequence;

[0160] A post-cleaning daily photovoltaic system efficiency prediction module is used to set the cleaning mark of the first second meteorological inverter data vector in the second meteorological inverter data vector sequence to clean, and set the cleaning marks of other second meteorological inverter data vectors to not clean, and input the second meteorological inverter data vector sequence with the updated cleaning mark into the first double-layer LSTM neural network model to predict the post-cleaning daily photovoltaic system efficiency for the next N days;

[0161] An uncleaned daily photovoltaic system efficiency prediction module is used to set the cleaning marks of all second meteorological inverter data vectors in the second meteorological inverter data vector sequence to not be cleaned, input the second meteorological inverter data vector sequence with the updated cleaning mark to the first double-layer LSTM neural network model, and predict the uncleaned daily photovoltaic system efficiency for the next N days;

[0162] A daily cleaning efficiency gain determination module is used to divide the daily photovoltaic system efficiency after cleaning for the next N days by the daily photovoltaic system efficiency before cleaning to obtain the daily cleaning efficiency gain for the next N days;

[0163] Among them, the daily forecast meteorological data includes: weather type, cumulative radiation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and / or air quality index; daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency.

[0164] Furthermore, the profit determination unit 203 is configured to: calculate the predicted power generation gain after cleaning according to formula (1); and calculate the predicted profit of cleaning on the first day of the next N days according to formula (2).

[0165] Furthermore, the device also includes: a second spatiotemporal graph neural network model training unit;

[0166] The second spatiotemporal graph neural network model training unit is configured to: obtain historical daily meteorological data, historical daily inverter data and historical daily measured power generation for multiple days; organize the historical daily meteorological data, historical daily inverter data and historical daily measured power generation for multiple days into a first training vector on a daily basis, and multiple first training vectors corresponding to multiple days constitute a first training set; wherein the historical daily meteorological data and historical daily inverter data are used as training data, and the historical daily measured power generation is used as label data; the first training set is input into the second spatiotemporal graph neural network model, and the historical daily predicted power generation of the corresponding day is predicted based on the historical daily meteorological data and the historical daily inverter data; a preset first loss function value is calculated based on the historical daily predicted power generation and the historical daily measured power generation, and the model parameters of the second spatiotemporal graph neural network model are updated based on the gradient descent method according to the first loss function value until a preset first convergence condition is reached, and the model parameters of the second spatiotemporal graph neural network model are updated using the obtained model parameters of the second spatiotemporal graph neural network model;

[0167] Among them, a second spatiotemporal graph neural network model is pre-constructed, and the model structure of the second spatiotemporal graph neural network model is the same as that of the first spatiotemporal graph neural network model; the historical daily meteorological data includes: weather type, radiation accumulation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and AQI; the historical daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency; the format of the first training vector is: Wherein, W is the weather type; R is the accumulated radiation, unit: megajoule / square meter; T is the temperature, unit: Celsius; H is the humidity, unit: %; D is the wind direction, unit: angle; V is the wind speed, meter / second; PM10 is the PM10 concentration, unit: %, PM2.5 is the PM2.5 content, unit: %; AQI is the air quality index, unit: microgram / cubic meter; are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter, G is the daily measured power generation, unit: kilowatt-hour.

[0168] Furthermore, the device also includes: a second LSTM neural network model training unit;

[0169] The second LSTM neural network model training unit is configured as follows:

[0170] Obtain multiple days of historical daily meteorological data, historical daily inverter data, historical daily cleaning data, and historical daily measured photovoltaic system efficiency;

[0171] The historical daily meteorological data, historical daily inverter data, historical daily cleaning and historical daily measured photovoltaic system efficiency of multiple days are used to form a second training vector on a daily basis, and multiple second training vectors corresponding to multiple days constitute a second training set; wherein the historical daily meteorological data, historical daily inverter data and historical daily cleaning are used as training data, and the historical daily measured photovoltaic system efficiency is used as label data;

[0172] The second training set is input into the second LSTM neural network model, and the photovoltaic system efficiency of the corresponding day is predicted based on the historical daily meteorological data, the historical daily inverter data and whether the historical daily cleaning data has been predicted;

[0173] According to the historical daily predicted photovoltaic system efficiency and the historical daily measured photovoltaic system efficiency, a preset second loss function value is calculated, and according to the second loss function value, the model parameters of the second LSTM neural network model are updated based on the gradient descent method until a preset second convergence condition is reached, and the model parameters of the second LSTM neural network model are used to update the model parameters of the first LSTM neural network model;

[0174] Among them, a second LSTM neural network model is pre-built, and the model structure of the second LSTM neural network model is the same as that of the first LSTM neural network model; the historical daily meteorological data includes: weather type, radiation accumulation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and AQI; the historical daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency; the format of the second training vector is:

[0175] Wherein, W is the weather type; R is the accumulated radiation, unit: megajoule / square meter; T is the temperature, unit: Celsius; H is the humidity, unit: %; D is the wind direction, unit: angle; V is the wind speed, meter / second; PM10 is the PM10 concentration, unit: %, PM2.5 is the PM2.5 content, unit: %; AQI is the air quality index, unit: microgram / cubic meter; are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter, C is whether it has been cleaned on that day, and PR is the daily measured photovoltaic system efficiency.

[0176] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0177] The above technical solution of the embodiment of the present invention is described in detail below in conjunction with specific application examples. For technical details not introduced during the implementation process, please refer to the relevant description in the previous text.

[0178] The embodiment of the present invention is mainly based on load forecasting and dust accumulation analysis, introduces the calculation of cleaning costs, and optimizes the cleaning plan. Load forecasting: generally refers to the prediction of the power generation of distributed photovoltaic power generation systems in the future. Since photovoltaics are most directly related to solar radiation intensity, a large amount of rich weather forecast data is involved. At present, the more mainstream load forecasting algorithms are mostly deep learning algorithms based on neural networks. Considering that distributed photovoltaic load forecasting may have highly complex spatiotemporal correlations, data from different sources have delays and other problems. The embodiment of the present invention adopts a spatiotemporal graph neural network model. Compared with the general neural network model, the spatiotemporal graph model can simultaneously derive time and space correlations, effectively reducing the problem of delays in data from different sources. Time correlation refers to the correlation of the observed values ​​of time series data in time, that is, the current value may be affected by past values. Spatial correlation refers to the correlation of data from adjacent geographical locations, and can also refer to the correlation between data from similar sources. In load forecasting, the spatiotemporal correlation between photovoltaic power generation and weather information is relatively complex. For example, if the radiation and power generation of the previous hour are known, it is impossible to calculate the power generation of the next hour through the radiation of the next hour and a simple mathematical formula. It is also impossible to simply modify the existing prediction function of a certain inverter for another inverter, although they are all intrinsically related. The delay of different source data refers to the data from the field collector and the data obtained from the weather forecast. There may be a delay in time and it is not possible to achieve a very accurate one-to-one correspondence at the same time. Especially for meteorological data, the accumulated radiation value of a certain hour may not be completely consistent with the actual situation. The complexity of spatiotemporal correlation and the delay of different source data have a certain negative impact on the prediction.

[0179] The embodiment of the present invention adopts the STGAT (Space-Time Graph Attention Network) model for load forecasting. Graph Neutral Network (GNN) is a deep learning method based on graph structure, which can process data formats represented by graph structure. In the graph, each node is defined by its own characteristics and its adjacent nodes and relationships. The network calculates the representation vector of the node by recursively aggregating and transforming the representation vectors of adjacent nodes. Graph neural network has a flexible structure and update method, which can well express the structural characteristics of some data itself. Figure 3That is, a simple schematic diagram of a graph neural network. GAT (Graph Attention Network) is an improvement on the general GNN model, adding an attention mechanism. The core working principle of GAT is to calculate the relationship between nodes through the attention mechanism. The GAT model usually consists of multiple attention heads, each of which can learn different attention weights, and finally aggregate the outputs of multiple heads. This multi-head attention mechanism helps to improve the representation and generalization capabilities of the model. For GAT, when aggregating neighbor information, the weight information of different neighbors is considered, and then the weight information of these neighbors is weighted according to the attention score.

[0180]

[0181] Among them, e ij represents the attention score between node i and node j, φ is the MLP (Multilayer Perceptron) operation, and II represents the execution of vector concatenation operation;

[0182] The graph attention mechanism is used to calculate the parameters of the "importance" of each node in the graph to other nodes, allowing nodes to add the characteristics of their neighbors. Here W represents the weight matrix of each graph node, W i is the weight matrix of node i, W j is the node j, W k It refers to each node when all adjacent nodes are summed, and N(i)∪i is all the adjacent nodes of node i; Compared with ReLU, the negative part of LeakyReLU is no longer directly 0, but multiplied by a very small slope k, which can solve the problem of ReLU neuron death because the negative part still has a small gradient; α is a weight vector used for parameterization.

[0183] Then information aggregation is performed based on the attention scores:

[0184] h′ i =e ii W i h i +∑ j∈N(i) e ij W j h j (4)

[0185] Among them, h i Refers to input, h' i It refers to output.

[0186] The Spatial-Temporal Graph Neutral Network (STGNN) is a graph neural network with time series, such as Figure 4 As shown in Figure 1, each time step in STGNN is a graph and is passed through the GCN / GAT network to obtain the resulting encoded graph that embeds the spatial interdependencies of the data. These encoded graphs can be modeled like time series data, and only the integrity of the graph structure of the data at each time step needs to be preserved. t ∈R N×F |t=0,…,T}, where N is the number of spatial nodes and F is the feature dimension. For such data, the space-time graph is an effective structure that can characterize the relationship between different nodes within a certain space-time range. The space-time graph can be expressed as: G t =(V,E t ,A t ), where V is the node set, E and A are the edge set and adjacency matrix. In general, the weather forecast data needs to support at least 14 days in the future, denoted as n, so that the forecast load data corresponding to n days can be generated respectively, that is, L = (L1, L2, ..., L n ), L n Represents the forecast load data for the nth day.

[0187] Analysis of dust accumulation: The dust accumulation degree of photovoltaic panels varies significantly according to the local air environment and climatic conditions. In the relatively humid environment of southern China, the dust loss over time is not obvious compared to the arid environment. At the same time, after cleaning, the low dust loss is maintained for a longer time. In arid environments such as northern China, dust on photovoltaic panels will continue to accumulate significantly even after a long period of time. Even if the dust has a considerable thickness, it will form a mechanical structure that tends to a steady state, rather than becoming thinner or being blown away by the wind as hoped. Cleaning or precipitation is the best way to reduce photovoltaic dust loss.

[0188] There is no accurate definition of the degree of dust accumulation in photovoltaics. The PR value can be used as a description of photovoltaic power generation efficiency. The PR value is the photovoltaic system efficiency (Performance Ratio), which is a key indicator for evaluating the quality of a photovoltaic system. It is the ratio of the actual output power of the power station to the theoretical output power, reflecting the proportional relationship of the actual input power to the grid after deducting all losses of the entire power station (including radiation loss, line loss, device loss, dust loss, heat loss, etc.). Its standardized calculation formula is generally carried out in accordance with the provisions of IEC61724.

[0189]

[0190] Among them, Eout is the power generation of the photovoltaic system in a specific period of time (unit: kilowatt-hour, kWh); P0 is the rated power of the photovoltaic system under STC (Standard Test Conditions) (unit: kilowatt, KW); H i is the radiation received by the photovoltaic array surface in a specific period of time (unit: kilowatt-hour per square meter, kWh / m2); G i The reference light intensity (1kW / m2) is used to define P0.

[0191] As the dust accumulation increases, the photovoltaic power generation efficiency gradually decreases, and its PR value will continue to decline. For such complex and long-term nonlinear sequence problems, the LTSM long short-term memory neural network is the first choice for prediction. Compared with the RNN long short-term memory network, the LSTM (Long Short Term Memory Network) long short-term memory neural network has the advantages of better information storage and more accurate access to historical information, and can avoid the gradient disappearance problem in the RNN network.

[0192] like Figure 5 As shown in the figure, it is the structure of the LTSM memory module. Now assume that an LSTM neural network input has I dimensions, the number of hidden layer nodes is H, and the number of output nodes is K. Then its forward propagation formula is formula (6) and formula (7):

[0193]

[0194]

[0195] Among them, w is the weight between nodes; is the input passing through node j at time i; is the output passing through node j at time i.

[0196] The formulas for back propagation are formula (8), formula (9) and formula (10):

[0197]

[0198]

[0199]

[0200] in, is the output error of the gate and the cell at time i.

[0201] The following is an explanation of the model establishment of the LTSM neural network for PR value prediction:

[0202] Input layer: Select influencing factors including but not limited to: light intensity, temperature, humidity, wind direction, wind speed, weather type, PM10 concentration, PM2.5 content, AQI (Air Quality Index), historical power generation, inverter voltage, inverter current, inverter efficiency, inverter temperature, photovoltaic panel temperature, and whether cleaning has occurred; the time of different input data corresponds to the output time, and on this basis, data cleaning, null value filling, normalization and other preprocessing operations are performed. Define N+1 data samples, the first N samples are influencing factors and historical power generation, used for training, and the last group of samples consists of influencing factors of the forecast day, which are used as test samples, i.e. forecast input. The PR value is input as a label, not a feature vector, and is used to calculate the loss value of the result. When used at the end, the power generation is the forecast result. First, use historical data to train the model. For example, use data from the past year at 1-hour intervals to train the model. Input groups of historical radiation accumulation, temperature, humidity and other weather data, as well as whether cleaning has occurred. The output results are used to calculate the loss function with the historical power generation power. Then, the parameters are corrected through back propagation. This process is repeated continuously. After the training is completed, I only need to input the weather data for each hour of the next n days into the model to get the forecast results for each hour of the next n days.

[0203] Hidden layer: For the learning ability of future information, a single LSTM can be added with a LSTM in the same direction to form two LSTM neural network structures to ensure the integrity of the prediction time before and after. The prediction model is as follows Figure 6 shown.

[0204] Output layer: After passing through the fully connected layer, the output result is the predicted PR value for the next n days. The PR before cleaning can be obtained by changing the parameter whether to clean it or not. c1 ,…,PR cn ) value and PRw after cleaning = (PR w1 ,…,PR wn ) value changes. Whether to clean is a feature vector used as input in training. Generally, 0 represents uncleaned and 1 represents cleaned, so that the PR value decay of cleaned and uncleaned can be predicted.

[0205] Then the ratio D can be calculated by formula (11): n To define the cleaning efficiency gain of photovoltaic panels, that is,

[0206]

[0207] The photovoltaic attenuation value caused by non-dust accumulation needs to be considered here. Experiments have shown that for a prediction time range of less than 30 days, the non-dust accumulation attenuation is almost negligible. However, since the weather forecast accuracy for more than 30 days is poor, it cannot be used as input data for the embodiments of the present invention. Therefore, the continuous attenuation of PR can be roughly used as an indicator of the degree of dust accumulation.

[0208] Cleaning plan: After having the predicted load (daily power generation) L and the cleaning gain degree D, the cleaning plan can be calculated. According to formula (1), the predicted power generation gain W after cleaning can be determined; where D n Indicates the cleaning efficiency gain on the nth day, L n represents the predicted load on the nth day. After cleaning, the power generation gain will generally decay over time, so n D will be generated, namely D1 to D n In actual production, a complex cost accounting method is generally used to calculate whether this benefit can meet the demand. Here is a simpler calculation method:

[0209] f(W)=charge(W)-C-φ (12)

[0210] Among them, income = charge (W) - C, W is the predicted power generation gain after cleaning, charge is the function of converting electricity into electricity charges, C is the cleaning cost of day t, and φ is the basic income. The charge function is the electricity price charging standard provided by the power plant to the owner, which is different in different regions and projects. The cleaning cost is the cost required for a cleaning (such as labor costs or transportation costs, etc.), and the basic income is a minimum income standard. If it is lower than the minimum income, there is no need to clean even if the income is greater than 0. Therefore, a cleaning plan can be determined, that is, cleaning is recommended when f9W)>0. At this time, the income meets the requirements and cleaning can be carried out on the same day.

[0211] The sources of the original data of the embodiment of the present invention are described below. The data mainly comes from the following three sources:

[0212] 1. From the site of the station, through the data collector (edge ​​gateway), directly connect to the on-site inverter, meter, weather station and other equipment, directly obtain the real-time data on the site, and send it to the database server. It is mainly the daily inverter power generation (kWh) and the cumulative radiation collected by the weather station (that is, the cumulative value of the radiation intensity, MJ / m2). 2. Data from meteorological forecasts generally need to be directly obtained from meteorological forecast-related agencies, and are divided into historical meteorological data and future forecast meteorological data. Mainly including daily weather types (such as sunny, cloudy, cloudy, etc.), cumulative radiation (here refers to the cumulative radiation obtained from the meteorological forecast, which is different from the radiation collected from the on-site meteorological station of the station, MJ / m2), temperature (℃), humidity (%), wind direction (angle), wind speed (m / s), PM10 concentration (%), PM2.5 content (%), AQI (Air Quality Index, Air Quality Index, μg / m3) indicators. 3. Manually entered data, such as whether it has been cleaned.

[0213] The following is an explanation of how to organize the input data. Generally, the time interval of time series data is one day. Figure 7 This is the daily power generation data of inverter No. N-3-1-1 of a station. Aggregate all the required data of each inverter together and sort them in chronological order. The data organization form for load forecasting is as follows:

[0214] W is the weather type, such as sunny, cloudy, overcast, etc. When used as input data, it needs to be converted into a floating point number, such as sunny = 1, cloudy = 2, etc. R, T, H, D, V, PM10, PM2.5, AQI correspond to the cumulative radiation (MJ / m2), temperature (℃), humidity (%), wind direction (angle), wind speed (m / s), PM10 concentration (%), PM2.5 content (%), AQI (Air Quality Index, Air Quality Index, μg / m3) indicators, respectively. are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter, and the above data are all used as feature vectors. G is the daily power generation of the inverter, which is used as the label of the training data to calculate the loss function value. After the model training is completed, the previous feature vector is input again to get the predicted daily power generation. Because the cleaning benefit analysis requires the power generation of the station, the daily power generation data of each inverter needs to be added up to get the daily power generation of the station. Data organization form for dust accumulation analysis: Where C is whether cleaning has been performed. It needs to be converted into a floating point number like the weather type. Generally, cleaning is 1 and uncleaning is 0. The others are similar to load prediction. PR (photovoltaic system efficiency) is used as a label to participate in the calculation of the loss function. After the model is trained, it is output as the prediction result. PR is calculated according to formula (5) through the daily power generation of the station and the accumulated radiation collected by the meteorological station in the station.

[0215] Beneficial effects: The embodiment of the present invention evaluates whether to clean based on the power generation income and cleaning cost, which is simple, convenient and low-cost. It effectively optimizes the judgment logic of when to clean the photovoltaic panels, provides the best cleaning plan, reduces unnecessary cleaning work, effectively avoids the additional costs and economic losses caused by poor cleaning effects due to seasons and weather, and maximizes the benefits of cleaning. The embodiment of the present invention can simply adapt data of different ranges to achieve a method for calculating cleaning benefits at different granularity levels. There is no need to modify the algorithm due to changes in the calculation range. For example, the current algorithm is based on days and stations. Under normal circumstances, the accuracy of cleaning benefit calculation in days and stations is sufficient; under other requirements, if a shorter time interval (such as hours or even minutes) is required to more accurately locate the dust accumulation (for example, specific to the roof or photovoltaic panel), there is no need to significantly change the embodiment of the present invention. It is only necessary to change the format of the input data and the dimension of the coefficient tensor in the model.

[0216] It should be understood that the specific order or hierarchy of steps in the disclosed process is an example of an exemplary method. Based on design preferences, it should be understood that the specific order or hierarchy of steps in the process can be rearranged without departing from the scope of protection of the present disclosure. The attached method claims present the elements of the various steps in an exemplary order and are not intended to be limited to the specific order or hierarchy described.

[0217] In the above detailed description, various features are grouped together in a single embodiment to simplify the disclosure. This method of disclosure should not be interpreted as reflecting an intention that the embodiments of the claimed subject matter require more features than are clearly stated in each claim. On the contrary, as reflected in the appended claims, the invention is in a state of having less than all the features of the disclosed individual embodiments. Therefore, the appended claims are hereby expressly incorporated into the detailed description, with each claim standing on its own as a separate preferred embodiment of the invention.

[0218] The disclosed embodiments are described above to enable any person skilled in the art to implement or use the present invention. Various modifications of these embodiments are obvious to those skilled in the art, and the general principles defined herein may also be applied to other embodiments without departing from the spirit and scope of the present disclosure. Therefore, the present disclosure is not limited to the embodiments given herein, but is consistent with the broadest scope of the principles and novel features disclosed in this application.

[0219] The above description includes examples of one or more embodiments. Of course, it is impossible to describe all possible combinations of components or methods for the purpose of describing the above embodiments, but it should be recognized by those skilled in the art that the various embodiments may be further combined and arranged. Therefore, the embodiments described herein are intended to cover all such changes, modifications and variations that fall within the scope of protection of the appended claims. In addition, with respect to the term "comprising" used in the specification or claims, the word is covered in a manner similar to the term "including". In addition, any term "or" used in the specification of the claims is intended to mean "non-exclusive or".

[0220] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A cleaning benefit prediction method based on load prediction and dust accumulation analysis, characterized in that: include: Obtain daily forecast weather data for the next N days from weather forecast agencies, and obtain daily inverter data for the recent N days from stations; According to the daily forecast meteorological data for the next N days and the daily inverter data for the recent N days, based on the first spatiotemporal graph neural network model, predict the daily forecast power generation for the next N days; According to the daily forecast meteorological data for the next N days and the daily inverter data for the recent N days, based on the first double-layer LSTM neural network model, predict the daily cleaning efficiency gain for the next N days; Based on the daily predicted power generation, daily cleaning efficiency gain, and cleaning cost for the next N days, the predicted benefits of cleaning on the first day of the next N days are predicted; Determine whether the predicted return of cleaning on the first day of the next N days is greater than the basic return; If the predicted return of cleaning on the first day of the next N days is greater than the basic return, cleaning will be performed on the first day of the next N days, otherwise no cleaning will be performed; Among them, the daily cleaning efficiency gain represents the ratio of the daily predicted photovoltaic system efficiency after cleaning to the daily predicted photovoltaic system efficiency before cleaning; the daily predicted photovoltaic system efficiency after cleaning represents the daily predicted photovoltaic system efficiency for the next N days if cleaning is performed on the first day of the next N days; the daily predicted photovoltaic system efficiency before cleaning represents the daily predicted photovoltaic system efficiency for the next N days if no cleaning is performed in the next N days; the network parameters of the first spatiotemporal graph neural network model are pre-trained based on historical daily meteorological data, historical daily inverter data and historical daily power generation; the network parameters of the first two-layer LSTM neural network model are pre-trained based on historical daily meteorological data, historical daily inverter data, historical daily cleaning conditions and historical daily photovoltaic system efficiency.

2. The cleaning benefit prediction method based on load prediction and dust accumulation analysis as claimed in claim 1, characterized in that: The method predicts the daily predicted power generation for the next N days based on the daily predicted meteorological data for the next N days and the daily inverter data for the recent N days in history, based on the first spatiotemporal graph neural network model, including: Encapsulating the daily forecast meteorological data for the next N days and the daily inverter data for the most recent N historical days into N first meteorological inverter data vectors, and sorting them in chronological order to obtain a first meteorological inverter data vector sequence; Inputting the first meteorological inverter data vector sequence into a first spatiotemporal graph neural network model to predict daily power generation for the next N days; Among them, the daily forecast meteorological data includes: weather type, cumulative radiation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and / or air quality index; daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency.

3. The cleaning benefit prediction method based on load prediction and dust accumulation analysis according to claim 1, characterized in that: The method predicts the daily cleaning efficiency gain for the next N days based on the daily forecasted meteorological data for the next N days and the daily inverter data for the recent N days in history, based on the first double-layer LSTM neural network model, including: Encapsulating the daily forecast meteorological data for the next N days, the daily cleaning mark, and the daily inverter data for the most recent N historical days into N second meteorological inverter data vectors, and sorting them in chronological order to obtain a second meteorological inverter data vector sequence; The cleaning mark of the first second meteorological inverter data vector in the second meteorological inverter data vector sequence is set to be cleaned, and the cleaning marks of other second meteorological inverter data vectors are set to be not cleaned, and the second meteorological inverter data vector sequence with updated cleaning marks is input into the first double-layer LSTM neural network model to predict the daily photovoltaic system efficiency after cleaning for the next N days; Setting the cleaning marks of all second meteorological inverter data vectors in the second meteorological inverter data vector sequence to not being cleaned, inputting the second meteorological inverter data vector sequence with the updated cleaning marks into the first double-layer LSTM neural network model, and predicting the un-cleaned daily photovoltaic system efficiency for the next N days; Divide the daily PV system efficiency after cleaning by the daily PV system efficiency before cleaning for the next N days to obtain the daily cleaning efficiency gain for the next N days; Among them, the daily forecast meteorological data includes: weather type, cumulative radiation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and / or air quality index; daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency.

4. The cleaning benefit prediction method based on load prediction and dust accumulation analysis according to claim 1, characterized in that: The predicted benefit of cleaning on the first day of the next N days is predicted based on the daily predicted power generation, daily cleaning efficiency gain, and cleaning cost of the next N days, including: The predicted post-cleaning power generation gain is calculated according to the following formula: <h2 style=";text-align:left;direction:ltr">W=D1L1+…+D<h2 style=";text-align:left;direction:ltr"> n <h2 style=";text-align:left;direction:ltr"> L<h2 style=";text-align:left;direction:ltr"> n Where W is the predicted power generation gain after cleaning, Dn is the daily cleaning efficiency gain on the nth day in the future, and Ln is the daily predicted power generation on the nth day in the future; The predicted return on cleaning on the first day of the next N days is calculated according to the following formula: income=charge(W)-C Among them, income is the predicted income of cleaning on the first day of the next N days, charge() is the electricity price charging function provided by the plant to the owner, and C is the cleaning cost of one cleaning.

5. The cleaning benefit prediction method based on load prediction and dust accumulation analysis according to claim 1, characterized in that: The method further comprises: Obtain historical daily meteorological data, historical daily inverter data, and historical daily measured power generation for multiple days; The historical daily meteorological data, the historical daily inverter data and the historical daily measured power generation of multiple days are used to form a first training vector on a daily basis, and multiple first training vectors corresponding to multiple days constitute a first training set; wherein the historical daily meteorological data and the historical daily inverter data are used as training data, and the historical daily measured power generation is used as label data; Inputting the first training set into a second spatiotemporal graph neural network model, and predicting the historical daily predicted power generation of the corresponding day according to the historical daily meteorological data and the historical daily inverter data; According to the historical daily predicted power generation and the historical daily measured power generation, a preset first loss function value is calculated, and according to the first loss function value, the model parameters of the second spatiotemporal graph neural network model are updated based on the gradient descent method until a preset first convergence condition is reached, and the model parameters of the second spatiotemporal graph neural network model are used to update the model parameters of the first spatiotemporal graph neural network model; Wherein, a second spatiotemporal graph neural network model is pre-constructed, and the model structure of the second spatiotemporal graph neural network model is the same as that of the first spatiotemporal graph neural network model; The historical daily meteorological data includes: weather type, radiation accumulation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and AQI; the historical daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency; The format of the first training vector is: Wherein, W is the weather type; R is the accumulated radiation, unit: megajoule / square meter; T is the temperature, unit: Celsius; H is the humidity, unit: %; D is the wind direction, unit: angle; V is the wind speed, meter / second; PM10 is the PM10 concentration, unit: %, PM2.5 is the PM2.5 content, unit: %; AQI is the air quality index, unit: microgram / cubic meter; are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter, G is the daily measured power generation, unit: kilowatt-hour.

6. The cleaning benefit prediction method based on load prediction and dust accumulation analysis according to claim 1, characterized in that: The method further comprises: Obtain historical daily meteorological data, historical daily inverter data, historical daily cleaning data, and historical daily measured photovoltaic system efficiency for multiple days; The historical daily meteorological data, historical daily inverter data, historical daily cleaning and historical daily measured photovoltaic system efficiency of multiple days are used to form a second training vector on a daily basis, and multiple second training vectors corresponding to multiple days constitute a second training set; wherein the historical daily meteorological data, historical daily inverter data and historical daily cleaning are used as training data, and the historical daily measured photovoltaic system efficiency is used as label data; The second training set is input into the second LSTM neural network model, and the photovoltaic system efficiency of the corresponding day is predicted based on the historical daily meteorological data, the historical daily inverter data and whether the historical daily cleaning data has been predicted; According to the historical daily predicted photovoltaic system efficiency and the historical daily measured photovoltaic system efficiency, a preset second loss function value is calculated, and according to the second loss function value, the model parameters of the second LSTM neural network model are updated based on the gradient descent method until a preset second convergence condition is reached, and the model parameters of the second LSTM neural network model are used to update the model parameters of the first LSTM neural network model; Wherein, a second LSTM neural network model is pre-constructed, and the model structure of the second LSTM neural network model is the same as that of the first LSTM neural network model; The historical daily meteorological data includes: weather type, radiation accumulation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and AQI; the historical daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency; The format of the second training vector is: Wherein, W is the weather type; R is the accumulated radiation, unit: megajoule / square meter; T is the temperature, unit: Celsius; H is the humidity, unit: %; D is the wind direction, unit: angle; V is the wind speed, meter / second; PM10 is the PM10 concentration, unit: %, PM2.5 is the PM2.5 content, unit: %; AQI is the air quality index, unit: microgram / cubic meter; are the average three-phase voltage of the inverter, in kV, are the average three-phase current of the inverter, in ampere, is the average temperature of the inverter, unit: Celsius, is the average efficiency of the inverter, C is whether it has been cleaned on that day, and PR is the daily measured photovoltaic system efficiency.

7. A cleaning benefit prediction device based on load prediction and dust accumulation analysis, characterized in that: include: A data acquisition unit is used to acquire daily forecasted meteorological data for the next N days from a meteorological forecasting agency, and to acquire daily inverter data for the most recent N days from a station; A daily power generation prediction unit, configured to predict the daily predicted power generation for the next N days based on the daily predicted meteorological data for the next N days and the daily inverter data for the most recent N historical days, based on a first spatiotemporal graph neural network model; A daily cleaning efficiency gain prediction unit, configured to predict the daily cleaning efficiency gain for the next N days based on the daily predicted meteorological data for the next N days and the daily inverter data for the recent N historical days, based on a first double-layer LSTM neural network model; A revenue determination unit, for predicting a predicted revenue of cleaning on the first day of the next N days based on the predicted daily power generation, daily cleaning efficiency gain, and cleaning cost of the next N days; The income comparison unit is used to determine whether the predicted income of cleaning on the first day of the next N days is greater than the basic income; a cleaning scheme determination unit, configured to perform cleaning on the first day of the next N days if the predicted return of cleaning on the first day of the next N days is greater than the basic return, and not perform cleaning otherwise; Among them, the daily cleaning efficiency gain represents the ratio of the daily predicted photovoltaic system efficiency after cleaning to the daily predicted photovoltaic system efficiency before cleaning; the daily predicted photovoltaic system efficiency after cleaning represents the daily predicted photovoltaic system efficiency for the next N days if cleaning is performed on the first day of the next N days; the daily predicted photovoltaic system efficiency before cleaning represents the daily predicted photovoltaic system efficiency for the next N days if no cleaning is performed in the next N days; the network parameters of the first spatiotemporal graph neural network model are pre-trained based on historical daily meteorological data, historical daily inverter data and historical daily power generation; the network parameters of the first two-layer LSTM neural network model are pre-trained based on historical daily meteorological data, historical daily inverter data, historical daily cleaning conditions and historical daily photovoltaic system efficiency.

8. The cleaning benefit prediction device based on load prediction and dust accumulation analysis as claimed in claim 7, characterized in that: The daily power generation prediction unit is configured as follows: Encapsulating the daily forecast meteorological data for the next N days and the daily inverter data for the most recent N historical days into N first meteorological inverter data vectors, and sorting them in chronological order to obtain a first meteorological inverter data vector sequence; inputting the first meteorological inverter data vector sequence into a first spatiotemporal graph neural network model to predict the daily power generation for the next N days; Among them, the daily forecast meteorological data includes: weather type, cumulative radiation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and / or air quality index; daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency.

9. The cleaning benefit prediction device based on load prediction and dust accumulation analysis according to claim 7, characterized in that: The daily cleaning efficiency gain prediction unit comprises: A data encapsulation module, used for encapsulating the daily forecast meteorological data for the next N days, the daily cleaning mark and the daily inverter data for the recent N historical days into N second meteorological inverter data vectors, and sorting them in chronological order to obtain a second meteorological inverter data vector sequence; A post-cleaning daily photovoltaic system efficiency prediction module is used to set the cleaning mark of the first second meteorological inverter data vector in the second meteorological inverter data vector sequence to clean, and set the cleaning marks of other second meteorological inverter data vectors to not clean, and input the second meteorological inverter data vector sequence with the updated cleaning mark into the first double-layer LSTM neural network model to predict the post-cleaning daily photovoltaic system efficiency for the next N days; An uncleaned daily photovoltaic system efficiency prediction module is used to set the cleaning marks of all second meteorological inverter data vectors in the second meteorological inverter data vector sequence to not be cleaned, input the second meteorological inverter data vector sequence with the updated cleaning mark to the first double-layer LSTM neural network model, and predict the uncleaned daily photovoltaic system efficiency for the next N days; A daily cleaning efficiency gain determination module is used to divide the daily photovoltaic system efficiency after cleaning for the next N days by the daily photovoltaic system efficiency before cleaning to obtain the daily cleaning efficiency gain for the next N days; Among them, the daily forecast meteorological data includes: weather type, cumulative radiation, temperature, humidity, wind direction, wind speed, PM10 concentration, PM2.5 content and / or air quality index; daily inverter data includes average three-phase voltage, average three-phase current, average temperature and average efficiency.

10. The cleaning benefit prediction device based on load prediction and dust accumulation analysis according to claim 7, characterized in that: The income determination unit includes: The post-cleaning power generation gain determination module is used to calculate the predicted post-cleaning power generation gain according to the following formula: <h2 style=";text-align:left;direction:ltr">W=D1L1+…+D<h2 style=";text-align:left;direction:ltr"> n <h2 style=";text-align:left;direction:ltr"> L<h2 style=";text-align:left;direction:ltr"> n Where W is the predicted power generation gain after cleaning, Dn is the daily cleaning efficiency gain on the nth day in the future, and L n is the daily predicted power generation for the nth day in the future; The profit determination module is used to calculate the predicted profit of cleaning on the first day of the next N days according to the following formula: income=charge(W)-C Among them, income is the predicted income of cleaning on the first day of the next N days, charge() is the electricity price charging function provided by the plant to the owner, and C is the cleaning cost of one cleaning.

Images