A method and system for predicting clean energy power generation

Through real-time monitoring and multi-layer neural network prediction, the data quality and accuracy problems of clean energy generation power output prediction are solved, and the safety and stability of the power grid and energy utilization efficiency are improved.

CN114021822BActive Publication Date: 2024-11-05莫毓昌
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111317305.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-09
Publication Date
2024-11-05
Estimated Expiration
2041-11-09

AI Technical Summary

Technical Problem

The existing clean energy power generation power prediction methods have shortcomings in data quality and model accuracy, especially the randomness and volatility of the prediction of photovoltaic and wind power power outputs with strong environmental correlation, which affects the safety and stability of the power grid.

Method used

By monitoring the power output and working condition characteristic data of clean energy power generation equipment in real time, a high-quality training sample set is built, multi-layer neural network is used for prediction, and the prediction accuracy is improved by using near-neighbor and approximate data sequence prediction criteria, and alarms are generated when necessary information.

Benefits of technology

It improves the prediction accuracy of clean energy power generation power output, enhances the safety and stability of the power grid and energy utilization efficiency, reduces the calculation needs of the prediction process, and improves the prediction speed.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114021822B_ABST
    Figure CN114021822B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of clean energy utilization, and discloses a clean energy power generation prediction method and system. The clean energy power generation prediction method comprises the following steps: S1, real-time monitoring and data collection of clean energy power generation equipment operating condition data, and real-time monitoring and collection of power output data of the power generation equipment through an electric energy meter installed in the clean energy power generation equipment; S2, constructing high-quality data samples from historical monitoring data sequences for training power output machine learning models. The present invention provides a clean energy power generation prediction method and system, which can extract high-quality training samples, eliminate abnormal data sample interference, and adopt corresponding prediction means according to the operating condition characteristics of different situations to improve prediction accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of clean energy utilization, and in particular to a clean energy power generation prediction method and system. Background Art

[0002] As energy security and environmental protection have become the focus of attention in today's world, my country, as a major energy-consuming country, has clearly put forward the strategic goals of "carbon peak" before 2030 and "carbon neutrality" before 2060.

[0003] In the energy network, by integrating clean energy such as photovoltaics and wind power, a comprehensive energy system is constructed. The mutual coupling of photovoltaics, wind power, and coal power is achieved, the advantages of different energy systems are brought into play, and coordinated operation is carried out to improve energy utilization efficiency and the configuration and consumption of clean energy, so as to achieve energy security and environmental protection.

[0004] The electrical energy output of clean energy such as photovoltaics and wind power has a strong environmental correlation. For example, photovoltaic power output is closely related to factors such as irradiance, temperature, humidity, season, and weather type; wind power output is closely related to factors such as air velocity (wind speed), air flow direction (wind direction), air density, air pressure, and temperature.

[0005] Since clean energy power output is random, volatile and uncontrollable, when large-scale clean energy power generation equipment is connected to the power grid, it will have an adverse impact on the safety, stability and economic operation of the power system. Therefore, clean energy power output prediction is of great significance to power grid management, dispatching, operation, system optimization, full utilization of energy, and safe and stable operation of the power grid.

[0006] Currently, clean energy power generation prediction methods can be divided into two categories: physical model methods and machine learning methods.

[0007] For the physical model method, when building the model, it is necessary to have sufficient working environment and equipment manufacturing information to establish an accurate physical model for power prediction. However, the limitation of the physical method is that the information provided by the manufacturer is usually limited and only valid under normal working conditions. In addition, due to the limitations of technical means and capital investment, the working environment parameters of power generation equipment that can be collected are usually limited.

[0008] For machine learning methods, the operating data of clean energy power generation equipment is modeled and analyzed to find out the internal rules and analyze the relationship between power generation and operating characteristics for prediction. The main methods include artificial neural network (ANN) and support vector machine. These methods have achieved good results to a certain extent, but artificial neural network has the disadvantage of high quality requirements for training samples; support vector machine has the problem of difficulty in selecting penalty function, kernel function and kernel parameters. Summary of the invention

[0009] Based on this, the present invention provides a clean energy power generation prediction method and system, which can extract high-quality training samples, eliminate interference from abnormal data samples, and take corresponding prediction measures according to the operating characteristics of different situations to improve prediction accuracy.

[0010] In order to achieve the above object, the present invention adopts such technical solution:

[0011] A clean energy power generation prediction method comprises the following steps:

[0012] S1. Real-time monitoring and data collection of the operating data of the clean energy power generation equipment are performed. The power output data of the power generation equipment is monitored and collected in real time through the electric energy meter installed in the clean energy power generation equipment. The power output data is recorded as P(t); {P(1), P (2), ..., P(t)} is the power output historical monitoring data sequence; various sensors installed in the working environment of the clean energy power generation equipment are used to monitor the operating characteristics of the power generation equipment in real time and collect the operating characteristic data. The operating characteristic data is recorded as X1(t), X2(t), ..., XM(t), where M is the number of operating characteristics that can be collected, and {XM (1), XM (2), ..., XM (t)} is the operating characteristic XM historical monitoring data sequence;

[0013] S2. Construct high-quality data samples from the historical monitoring data sequence for training the power output machine learning model, which includes obtaining a high-quality training sample set, building a multi-layer neural network, and training the power output machine learning model. The specific operations are as follows:

[0014] S21. Obtain high-quality training sample set:

[0015] For any operating data sequence X(h) = {X1(h), X2(h), …, XM(h)}, h = 1, 2, …, t; find the operating data sequence X(w) adjacent to X(h), so-called adjacent satisfies the following conditions:

[0016] For any q = 1, 2, …, M, | Xq(h) - Xq(w) | < STDX(Q) / R

[0017] Among them, STDX(Q) is the sample standard deviation of the historical monitoring sequence of the operating condition feature Xq,

[0018] R is the threshold set by the user;

[0019] If the number of adjacent monitoring data sequences found, U, is greater than K,

[0020] K is the threshold set by the user;

[0021] Then the power output data sequence corresponding to these U working condition data sequences is recorded as, P(h) ={P(h1), P(h2),…,P(hu)}

[0022] If | min {P(h)} – max {P(h)} | < STDP / R, where STDP is the sample standard deviation of P(h);

[0023] Then the mean value MEANP(h) of the operating condition data sequence X(h) and the power output data sequence P(h) is taken as a high-quality training sample {X(h), MEANP(h)} and put into the high-quality training sample set;

[0024] S22. Constructing a multi-layer neural network:

[0025] The operating condition data of the high-quality training sample set is used as input, and the power output data of the high-quality training sample set is used as output to construct a multi-layer neural network;

[0026] The multi-layer neural network consists of an input layer, two hidden layers, and an output layer;

[0027] The input layer of the multivariate neural network includes M inputs, which come from the M operating condition features X1, ..., XM of the training samples;

[0028] The first hidden layer of the multivariate neural network includes P neurons, and the calculation formula is,

[0029] ;

[0030] in, It is the input of the multi-layer neural network; is the weight matrix; is bias; is the P-dimensional output vector of the first hidden layer;

[0031] The second hidden layer of the multivariate neural network includes Q neurons, and the calculation formula is,

[0032] ;

[0033] in, is the weight matrix; is bias; is the Q-dimensional output vector of the second hidden layer;

[0034] The output layer of the multivariate neural network includes 1 neuron, and the calculation formula is,

[0035] ;

[0036] in, is the weight matrix; is bias; is the output vector of the output layer, which includes 1 output, which corresponds to the power output of the training sample;

[0037] S23. Training power output machine learning model:

[0038] Use the gradient descent neural network training algorithm to train the multi-layer neural network and update the parameters of the multi-layer neural network. The specific steps include:

[0039] S231, network weights are initialized;

[0040] S232, the input data is forward propagated through two hidden layers to obtain an output value;

[0041] S233, finding the error between the output value of the network and the target value;

[0042] S234, when the error is greater than a given threshold value D, the error is transmitted back to the network, and the errors of the two hidden layers are obtained in sequence; when the error is equal to or less than the given threshold value D, the training is terminated;

[0043] S235, update the weights according to the obtained error, and then return to step S232;

[0044] S3. For the working condition characteristic data X1(T), X2(T), ..., XM(T) collected by various sensors currently installed in the working environment of the clean energy power generation equipment, the power output P*(T) of the current power generation equipment is predicted. The specific operation is as follows:

[0045] S31. In the training set constructed in step S21, the method for searching adjacent operating condition data sequences in step S21 is used to search for adjacent operating condition data sequences X(w) of the current operating condition data sequence X(T)={X1(T), X2(T), ..., XM(T)}. The following three processes are divided according to different search results:

[0046] S311. If the number of adjacent operating condition data sequences U of the current operating condition data sequence X(T) found in the training set constructed in step S21 is greater than 2, the adjacent operating condition data sequence X(w') closest to the current operating condition data sequence X(T) is selected, and the power output P(w') corresponding to the operating condition data sequence X(w') is used as the predicted value P*(T) of the power output of the current power generation equipment;

[0047] S312. If no operating condition data sequence adjacent to X(T) is found in the training set constructed in step S21, then in the historical monitoring data sequence, trace back a length of L from the current T-th operating condition data: XL(T)={X(TL), X(TL-2), ..., X(T-2), X(T-1)}, and its corresponding power output data sequence PL(T)={P(TL), P(TL-2), ..., P(T-2), P(T-1)}, and in the historical monitoring data sequence, search for an operating condition data sequence XL(t) similar to the current operating condition data sequence XL(T), where XL(t)={X(t-L), X(t-L-2), ..., X(t-2), X(t-1)};

[0048] The operating data sequences XL(t) and XL(T) are similar, and the conditions that need to be met are: any element X(t-i) in XL(t) and the corresponding element X(Ti) in XL(T) are adjacent, and the method of adjacent judgment uses the method of finding adjacent operating data sequences in step S21; the first similar operating data sequence found is recorded as XL(t*), and XL(t*) is the operating data sequence closest to the current time point T; the power output P(t*) corresponding to the operating data sequence X(t*) is used as the predicted value P*(T) of the power output of the current power generation equipment;

[0049] S313: No operating condition data sequence similar to the current operating condition data sequence XL(T) is found in the historical monitoring data sequence. At this time, the power output machine learning model constructed in step S2 can be used to output the predicted value. The specific operation is as follows:

[0050] Input the current operating condition data sequence X(T)={X1(T), X2(T), …, XM(T)} into the trained multi-layer neural network, and obtain the predicted value P*(T) of the current power output of the power generation equipment through the calculation of the neural network;

[0051] S4. According to the predicted value of the current power output of the power generating equipment outputted in step S3, abnormality detection and alarm can be performed. The specific operations are as follows:

[0052] According to the predicted value P*(T) output by the prediction module and the actual power output value P(T) monitored, the current monitored value is judged to be abnormal by whether it is a neighbor. If it satisfies

[0053] | P*(T) –P(T) | < STDP / R, then it is a nearest neighbor and no abnormality occurs;

[0054] Otherwise, it is judged that there is an abnormal hidden danger, and an alarm information SMS is generated and sent to the operation and maintenance management personnel, requesting the operation and maintenance management personnel to inspect the power generation equipment and eliminate the hidden danger of failure.

[0055] Further, in step S311, the distance between the current working condition data sequence X(T) and the adjacent working condition data sequence X(w) is calculated using the Euclidean distance calculation formula, wherein the distance calculation formula between the working condition data sequence X(T) and the adjacent working condition data sequence X(w') is as follows:

[0056] Deuclidean(X(T), X(w'))=min {Deuclidean(X(T), X(w)), operating condition data X(w) adjacent to X(T)}.

[0057] Furthermore, in step S1, for photovoltaic power generation equipment, the operating condition characteristic data include data on irradiance, temperature, humidity, season, and weather type at the work site; for wind power generation equipment, the operating condition characteristic data include data on air flow rate, air flow direction, air density, air pressure, and temperature at the work site.

[0058] The present application also provides a clean energy power generation power prediction system, which is characterized by comprising: an electric energy meter and sensor module, a historical monitoring database, a high-quality training set construction module, a power output neural network training module, a common operating condition feature prediction module, a high-dispersion operating condition feature prediction module, a special operating condition feature prediction module and a clean energy power generation anomaly detection module; wherein,

[0059] The electric energy meter and sensor module are used to monitor and collect power output data and operating condition characteristic data of the power generation equipment in real time, and send the monitoring data to the historical monitoring database;

[0060] The historical monitoring database is used to store the power output data and operating condition characteristic data sent by the electric energy meter and the sensor module;

[0061] The high-quality training set construction module is used to extract high-quality training samples from the historical monitoring data sequence stored in the historical monitoring database based on the aggregation degree of the data, eliminate the interference of abnormal data samples, and train the power output neural network model;

[0062] The power output neural network training module is used to train a multi-layer power output neural network based on a high-quality training set using a gradient descent neural network training algorithm, and fit the nonlinear relationship between power output and operating condition characteristics;

[0063] The common operating condition feature prediction module is used to search for adjacent operating condition data sequences from a high-quality training set using the nearest neighbor prediction criterion and output a predicted value using the statistical average output power;

[0064] The high-divergence operating condition characteristic prediction module is used to adopt the criterion of approximate data sequence prediction, calculate the similar operating condition data sequence matching the current operating condition data sequence from the historical monitoring data sequence, and use the power output of the most recent similar operating condition data sequence as the power output prediction value;

[0065] The special operating condition characteristic prediction module is used to adopt a neural network model that fits the nonlinear relationship between power output and operating condition characteristics, and output a predicted value through neural network forward propagation;

[0066] The clean energy power generation abnormality detection module determines whether the current detection value is abnormal based on the predicted value output by the prediction module and the actual power output value monitored. If it is abnormal, an alarm information SMS is generated and sent to the operation and maintenance management personnel.

[0067] At the same time, the present application provides an electronic device, characterized in that it includes a memory and a processor, and computer instructions stored in the memory and running on the processor, and when the computer instructions are run by the processor, the steps in the above method are executed.

[0068] In addition, the present application provides a computer-readable storage medium, characterized in that it is used to store computer instructions, and when the computer instructions are executed by a processor, the steps in the above method are executed.

[0069] Beneficial Effects

[0070] Compared with the prior art, the present invention has at least the following advantages:

[0071] 1. Improve the accuracy of power output prediction for clean energy sources with strong environmental relevance, such as photovoltaic power generation and wind power generation, so that clean energy such as photovoltaic power generation can be fully utilized and grid operation can be safer and more stable.

[0072] 2. Using the aggregation degree of data as a criterion, for any operating condition data sequence X(h), find the U operating condition data sequences X(w) adjacent to X(h) and the corresponding U power output data sequences P(h), take the mean MEANP(h) of the operating condition data sequence X(h) and the power output data sequence Ph as a valid data sample {X(h), MEANP(h)}, and put it into the training set, so as to extract high-quality training samples and eliminate the interference of abnormal data samples, thereby improving the prediction accuracy of the neural network.

[0073] 3. For operating characteristics with a high frequency of occurrence, the nearest neighbor prediction criterion is adopted. In the high-quality training sample set, the number of adjacent operating condition data sequences U of the current operating condition feature data X(T) is found to be greater than 2. Then the adjacent operating condition data sequence X(w') closest to the current operating condition data sequence X(T) is selected, and the power output P(w') corresponding to the operating condition data sequence X(w') is used as the predicted value P*(T) of the power output of the current power generation equipment, which can reduce the calculation requirements of the prediction process and improve the prediction speed.

[0074] 4. For scenarios where the operating condition data divergence is high due to the lack of core operating condition features, the machine learning model has the problem of poor accuracy. The approximate data sequence prediction criterion is used to search for the operating condition data sequence XL(t) similar to the current operating condition data sequence XL(T) in the historical monitoring data sequence. The first similar operating condition data sequence found is recorded as XL(t*), which is the operating condition data sequence closest to the current time point T. The power output P(t*) corresponding to the operating condition data sequence X(t*) is used as the predicted value P*(T) of the power output of the current power generation equipment. This can improve the prediction accuracy.

[0075] 5. For special operating conditions (i.e., operating conditions that rarely appear in history), the neural network model is used as the last resort for prediction. The current operating data sequence X(T) is input into the trained multi-layer neural network. After calculation by the neural network, the predicted value P*(T) of the current photovoltaic power generation equipment power output is obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION

[0077] 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.

[0078] See also Figure 1 , a clean energy power generation power prediction method, comprising the following steps:

[0079] S1. Real-time monitoring and data collection of operating data of clean energy power generation equipment such as photovoltaic power generation or wind power generation are carried out through electric energy meters and various sensors installed on the power generation equipment. Specifically, the power output data of the power generation equipment is monitored and collected in real time through electric energy meters installed in clean energy power generation equipment such as photovoltaic power generation or wind power generation. The power output data is recorded as P(t). {P(1), P (2), … , P(t)} is a sequence of historical monitoring data of power output; various sensors installed in the working environment of clean energy power generation equipment are used to monitor the operating characteristics of the power generation equipment in real time and collect operating characteristic data. The operating characteristic data are recorded as X1(t), X2(t), … , XM(t), where M is the number of operating characteristics that can be collected, among which {X1 (1), X1(2), … , X1(t)} is a sequence of historical monitoring data of operating characteristic X1, {X2 (1), X2 (2), … , X2 (t)} is a sequence of historical monitoring data of operating characteristic X2, and similarly, {XM (1), XM (2), … , XM (t)} is a sequence of historical monitoring data of operating characteristic XM; it should be noted that photovoltaic power generation is a technology that uses the photovoltaic effect of the semiconductor interface to directly convert light energy into electrical energy, and this light energy mainly comes from sunlight in nature. For photovoltaic power generation equipment, the working condition characteristic data collected by the sensor include data related to the environment such as irradiance, temperature, humidity, season, weather type, etc. at the working place; for wind power generation equipment, the working condition characteristic data collected by the sensor include data such as air flow rate, air flow direction, air density, air pressure, temperature, etc. at the working place. This embodiment is described in detail using photovoltaic power generation equipment.

[0080] S2. Construct high-quality data samples from the historical monitoring data sequence for training the power output machine learning model, which includes obtaining a high-quality training sample set, building a multi-layer neural network, and training the power output machine learning model. The specific operations are as follows:

[0081] S21. Obtain a high-quality training sample set, construct high-quality data samples, and eliminate the interference of abnormal data samples:

[0082] For any operating data sequence X(h) = {X1(h), X2(h), …, XM(h)}, h=1,2,…,t; for photovoltaic power generation equipment, in any operating data sequence X(h), M=5, X1(h) represents the irradiance data of sunlight, X2(h) represents the temperature data of the environment, X3(h) represents the environmental humidity data of the photovoltaic power generation equipment, X4(h) represents the seasonal parameter data, and X5(h) represents the weather type parameter data. Find the operating data sequence X(w) adjacent to X(h), the so-called adjacent satisfies the following conditions

[0083] For any q = 1, 2, …, M, | Xq(h) - Xq(w) | < STDX(Q) / R

[0084] Among them, STDX(Q) is the sample standard deviation of the historical monitoring sequence of the operating condition feature Xq,

[0085] R is the threshold set by the user, preferably, R=10;

[0086] If the number of adjacent monitoring data sequences found, U, is greater than K,

[0087] K is a threshold set by the user, preferably, K=10;

[0088] Then the power output data sequence corresponding to these U working condition data sequences is recorded as, P(h) ={P(h1), P(h2),…,P(hu)}

[0089] If | min {P(h)} – max {P(h)} | < STDP / R, where STDP is the sample standard deviation of P(h).

[0090] Then the mean value MEANP(h) of the operating condition data sequence X(h) and the power output data sequence P(h) is taken as a high-quality training sample {X(h), MEANP(h)} and put into the high-quality training sample set;

[0091] It should be noted that STDX(Q) is the sample standard deviation of the historical monitoring sequence Xq(1), Xq(2), … , Xq(t) of the operating condition feature Xq, MEANP(h) is the mean of the power output data sequence P(h1), P(h2), … , P(hu), and STDP is the sample standard deviation of the power output data sequence P(h1), P(h2), … , P(hu).

[0092] S22. Constructing a multi-layer neural network:

[0093] The operating condition data of the high-quality training sample set is used as input, and the power output data of the high-quality training sample set is used as output to construct a multi-layer neural network;

[0094] The multi-layer neural network consists of an input layer, two hidden layers, and an output layer;

[0095] The input layer of the multivariate neural network includes M inputs, which come from M operating condition features X1, ..., XM of the training samples respectively; in this embodiment, X1 is the illumination of the environment where the photovoltaic power generation equipment is located, X2 is the temperature of the environment where the photovoltaic power generation equipment is located, X3 is the ambient humidity of the photovoltaic power generation equipment, X4 is the season parameter, and X5 is the weather type parameter.

[0096] The first hidden layer of the multivariate neural network includes P neurons, and the calculation formula is,

[0097] ;

[0098] in, It is the input of the multi-layer neural network; is the weight matrix; is bias; is the P-dimensional output vector of the first hidden layer;

[0099] The second hidden layer of the multivariate neural network includes Q neurons, and the calculation formula is,

[0100] ;

[0101] in, is the weight matrix; is bias; is the Q-dimensional output vector of the second hidden layer;

[0102] The output layer of the multivariate neural network includes 1 neuron, and the calculation formula is,

[0103] ;

[0104] Among them, relu() is the relu activation function; is the weight matrix; is bias; is the output vector of the output layer, which includes 1 output, which corresponds to the power output of the training sample.

[0105] S23. Training power output machine learning model:

[0106] Use the gradient descent neural network training algorithm to train the multi-layer neural network and update the parameters of the multi-layer neural network. The specific steps include:

[0107] S231, network weights are initialized;

[0108] S232, the input data is forward propagated through two hidden layers to obtain an output value;

[0109] S233, finding the error between the output value of the network and the target value;

[0110] S234, when the error is greater than a given threshold D, the error is transmitted back to the network, and the errors of the two hidden layers are obtained in turn. When the error is equal to or less than the given threshold D, the training ends;

[0111] S235. Update the weights according to the obtained error, and then return to step S232.

[0112] S3. For various sensors currently installed in the working environment of photovoltaic clean energy power generation equipment, the collected working condition characteristic data X1(T), X2(T), ..., XM(T) are used to predict the power output P*(T) of the current photovoltaic power generation equipment. The specific operations are as follows:

[0113] S31. In the training set constructed in step S21, the method for searching for adjacent operating condition data sequences in step S21 is used to search for the adjacent operating condition data sequence X(w) of the current operating condition data sequence X(T)={X1(T), X2(T), ..., XM(T)}. Specifically, X(T) and X(W) are adjacent and meet the following conditions:

[0114] For any q=1, 2, …, M, | Xq(T) - Xq(w) | < STDX(Q) / R; where STDX(Q) is the sample standard deviation of the sequence Xq, and R is a threshold set by the user. In this implementation, R=10.

[0115] According to the different search results of the adjacent working condition data sequence X(w), it is divided into the following three processes. According to the different working condition characteristics, corresponding prediction methods are adopted to improve the prediction accuracy of the power output of photovoltaic power generation equipment.

[0116] S311. If the number U of adjacent operating condition data sequences of the current operating condition data sequence X(T) found in the training set constructed in step S21 is greater than 2, the current operating condition data sequence X(T) belongs to an operating condition with a higher frequency of occurrence. The neighbor prediction criterion is adopted to select the adjacent operating condition data sequence X(w') that is closest to the current operating condition data sequence X(T), and the power output P(w') corresponding to the operating condition data sequence X(w') is used as the predicted value P*(T) of the power output of the current power generation equipment; wherein, the distance between the current operating condition data sequence X(T) and the adjacent operating condition data sequence X(w) is calculated using the Euclidean distance calculation formula, and the operating condition data sequence X(w') has the smallest Euclidean distance among all the adjacent operating condition data found.

[0117] The distance calculation formula between the working condition data sequence X(T) and the adjacent working condition data sequence X(w') is as follows:

[0118] Deuclidean(X(T), X(w'))=min {Deuclidean(X(T), X(w)), working condition data X(w) adjacent to X(T)}; where Deuclidean() is the Euclidean distance calculation formula.

[0119] S312. If no operating condition data sequence adjacent to X(T) is found in the training set constructed in step S21, it indicates that the power output divergence of the current operating condition is large (the main reason is that the core operating condition feature of the power output is missing). In this case, the machine learning model has the problem of poor accuracy. At this time, the approximate data sequence prediction criterion is adopted. Specifically, in the historical monitoring data sequence, a working condition data sequence of length L is traced back from the current Tth working condition data: XL(T)={X(TL), X(TL-2),… , X(T-2), X(T-1)}, and its corresponding power output data sequence PL(T)={P(TL), P(TL-2),… , P(T-2), P(T-1)}.

[0120] In the historical monitoring data sequence, search for the operating condition data sequence XL(t) similar to the current operating condition data sequence XL(T), where XL(t)={X(t-L), X(t-L-2), …, X(t-2), X(t-1)};

[0121] The operating condition data sequences XL(t) and XL(T) are similar, and the condition to be satisfied is that any element X(t-i) in XL(t) and the corresponding element X(Ti) in XL(T) are adjacent. The method for determining the adjacent elements is the method for finding the adjacent operating condition data sequences in step S21, which is as follows:

[0122] A working condition data sequence XL(t) = {X(t-L), X(t-L-2), ..., X(t-2), X(t-1)} is similar to XL(T), and the condition that needs to be satisfied is that any element X(t-i) in XL(t) and the corresponding element X(Ti) in XL(T) are adjacent. X(t-i) and X(Ti) are adjacent if the following conditions are satisfied:

[0123] For any q=1, 2, … , M, | Xq(t-i)) - Xq(Ti) | < STDX(Q) / R, where STDX(Q) is the sample standard deviation of the sequence Xq, and R is a threshold set by the user. In this implementation, R=10.

[0124] Because the search is traced back from the current time point, the first similar operating condition data sequence found is recorded as XL(t*), and XL(t*) is the operating condition data sequence closest to the current time point T.

[0125] The power output P(t*) corresponding to the operating condition data sequence X(t*) is used as the predicted value P*(T) of the power output of the current photovoltaic power generation equipment.

[0126] S313: No operating condition data sequence similar to the current operating condition data sequence XL(T) is found in the historical monitoring data sequence, indicating that the current operating condition rarely occurs. At this time, the power output machine learning model constructed in step S2 can be used to output the predicted value. The specific operation is as follows:

[0127] Input the current operating condition data sequence X(T)={X1(T), X2(T), …, XM(T)} into the trained multi-layer neural network, and obtain the predicted value P*(T) of the current photovoltaic power generation equipment power output through the calculation of the neural network;

[0128] S4. According to the predicted value of the current photovoltaic power generation equipment power output output in step S3, abnormality detection and alarm can be performed. The specific operations are as follows:

[0129] According to the predicted value P*(T) output by the prediction module and the actual power output value P(T) monitored, the current monitored value is judged to be abnormal by whether it is a neighbor. If it satisfies

[0130] | P*(T) –P(T) | < STDP / R, where STDP is the sample standard deviation of the power output data sequence P(h1), P(h2), …, P(hu).

[0131] It is a neighbor, judging that no abnormality occurs;

[0132] Otherwise, it is judged that there is an abnormal hidden danger, and an alarm information SMS is generated and sent to the photovoltaic power generation equipment operation and maintenance management personnel, who will promptly feedback the alarm information and request the operation and maintenance management personnel to inspect the photovoltaic power generation equipment, eliminate the hidden dangers of failure, and ensure the safe operation of the photovoltaic power generation equipment.

[0133] The present invention also provides a clean energy power generation power prediction system, including: an electric energy meter and sensor module, a historical monitoring database, a high-quality training set construction module, a power output neural network training module, a common operating condition feature prediction module, a high-dispersion operating condition feature prediction module, a special operating condition feature prediction module and a clean energy power generation anomaly detection module; wherein,

[0134] The electric energy meter and sensor module are used to monitor and collect the power output data and operating condition characteristic data of the power generation equipment in real time, and send the monitoring data to the historical monitoring database. The operating condition characteristic data can be seen in the above text, and the operating condition characteristic data is recorded as X1(t), X2(t), ..., XM(t), where M is the number of operating condition characteristics that can be collected, which will not be repeated here.

[0135] The historical monitoring database is used to store the power output data and operating condition characteristic data sent by the electric energy meter and the sensor module.

[0136] The high-quality training set construction module is used to extract high-quality training samples from the historical monitoring data sequence stored in the historical monitoring database based on the degree of data aggregation, eliminate abnormal data sample interference, and train the power output neural network model. The method steps for extracting high-quality training samples and training the power output neural network model can be seen in the method of step S2 described above in the text, and will not be repeated here.

[0137] The power output neural network training module is used to train a multi-layer power output neural network based on a high-quality training set using a gradient descent neural network training algorithm, and fit the nonlinear relationship between power output and operating condition characteristics. The method steps for training a multi-layer power output neural network can be seen in the method of step S23 described above in the text, and will not be repeated here.

[0138] The common operating condition feature prediction module is used to search for adjacent operating condition data sequences from a high-quality training set based on the criterion of nearest neighbor prediction, and output a predicted value using the statistical average output power; the method for searching the adjacent operating condition data sequence can be seen in the method of step S31 described in the text above, which will not be repeated here.

[0139] The high-divergence operating condition characteristic prediction module is used to adopt the criterion of approximate data sequence prediction, calculate the similar operating condition data sequence that matches the current operating condition data sequence from the historical monitoring data sequence, and use the power output of the most recent similar operating condition data sequence as the power output prediction value; the method for finding the similar operating condition data sequence can be seen in the method of step S31 recorded in the text above, which will not be repeated here.

[0140] The special operating condition characteristic prediction module is used to adopt a neural network model that fits the nonlinear relationship between power output and operating condition characteristics, and output a predicted value through neural network forward propagation.

[0141] The clean energy power generation abnormality detection module determines whether the current detection value is abnormal based on the predicted value output by the prediction module and the actual power output value monitored. If abnormal, an alarm message is generated and sent to the operation and maintenance manager. The method for determining whether the current detection value is abnormal can be seen in the method of step S4 described above, which will not be repeated here.

[0142] At the same time, the present invention includes an electronic device, characterized in that it includes a memory and a processor and computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, all or part of the above steps are completed to realize the prediction of clean energy power generation such as photovoltaic power generation.

[0143] In addition, the present invention also includes a computer-readable storage medium, characterized in that it is used to store computer instructions, and when the computer instructions are executed by a processor, the steps in the above method are executed to complete all or part of the above steps. The computer-readable storage medium of the present invention may include but is not limited to a magnetic storage device, an optical disk, a digital versatile disk, a smart card, and a flash memory device.

[0144] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A clean energy power generation prediction method, characterized in that: It includes the following steps: S1. Monitor and collect the working condition data of clean energy power generation equipment in real time. Through the watt-hour meter installed in the clean energy power generation equipment, monitor and collect the power output data of the power generation equipment in real time, and record the power output data as P(t); {P(1), P(2), …, P(t)} is the historical monitoring data sequence of power output; through various sensors installed in the working environment of the clean energy power generation equipment, monitor the working condition characteristics of the power generation equipment in real time and collect the working condition characteristic data, and record the working condition characteristic data as X1(t), X2(t), …, XM(t), where M is the number of working condition characteristics that can be collected. Among them, {XM(1), XM(2), …, XM(t)} is the historical monitoring data sequence of the working condition characteristic XM; for photovoltaic power generation equipment, the working condition characteristic data includes the data of irradiance, temperature, humidity, season, and weather type at the working location; for wind power generation equipment, the working condition characteristic data includes the data of air velocity, air flow direction, air density, air pressure, and temperature at the working location. S2. Construct high-quality data samples from the historical monitoring data sequence for training the power output machine learning model, which includes obtaining a high-quality training sample set, constructing a multi-layer neural network, and training the power output machine learning model. The specific operations are as follows: S21. Obtain a high-quality training sample set: For any working condition data sequence X(h) = {X1(h), X2(h), …, XM(h)}, h = 1, 2, …, t; find the working condition data sequence X(w) adjacent to X(h), and the so-called adjacency satisfies the following conditions: For any q = 1, 2, …, M, |Xq(h) - Xq(w)| < STDX(Q) / R where STDX(Q) is the sample standard deviation of the historical monitoring sequence of the working condition characteristic Xq, and R is a threshold set by the user; If the number U of adjacent monitoring data sequences found is greater than K, where K is a threshold set by the user; then the power output data sequence corresponding to these U working condition data sequences is denoted as P(h) = {P(h1), P(h2), …, P(hu)} If |min{P(h)} – max{P(h)}| < STDP / R, where STDP is the sample standard deviation of P(h); then the mean value MEANP(h) of the working condition data sequence X(h) and the power output data sequence P(h) is used as a high-quality training sample {X(h), MEANP(h)} and put into the high-quality training sample set; S22. Construct a multi-layer neural network: Use the working condition data of the high-quality training sample set as the input and the power output data of the high-quality training sample set as the output to construct a multi-layer neural network; The multi-layer neural network includes an input layer, two hidden layers, and an output layer; The input layer of the multi-layer neural network includes M inputs, which come from the M working condition characteristics X1, …, XM of the training samples respectively; The first hidden layer of the multi-layer neural network includes P neurons, and the calculation formula is, O1 = relu(w1·x + b1); Among them, x is the input of the multi-layer neural network; w1 is the weight matrix; b1 is the bias; O1 is the output vector of the first hidden layer P dimension; The second hidden layer of the multivariate neural network includes Q neurons, and the calculation formula is, O2 = relu(w2·O1+b2); Among them, w2 is the weight matrix; b2 is the bias; O2 is the output vector of the second hidden layer Q dimension; The output layer of the multivariate neural network includes 1 neuron, and the calculation formula is, y=relu(w3·O2+b3) Among them, w3 is the weight matrix; b3 is the bias; y is the output vector of the output layer, which includes 1 output, which corresponds to the power output of the training sample; S23. Training power output machine learning model: Use the gradient descent neural network training algorithm to train the multi-layer neural network and update the parameters of the multi-layer neural network. The specific steps include: S231, network weights are initialized; S232, the input data is forward propagated through two hidden layers to obtain an output value; S233, finding the error between the output value of the network and the target value; S234, when the error is greater than a given threshold value D, the error is transmitted back to the network, and the errors of the two hidden layers are obtained in sequence; when the error is equal to or less than the given threshold value D, the training is terminated; S235, update the weights according to the obtained error, and then return to step S232; S3. For the working condition characteristic data X1(T), X2(T), ..., XM(T) collected by various sensors currently installed in the working environment of the clean energy power generation equipment, the power output P*(T) of the current power generation equipment is predicted. The specific operation is as follows: S31. In the training set constructed in step S21, the method for searching adjacent working condition data sequences in step S21 is used to search for adjacent working condition data sequences X(w) of the current working condition data sequence X(T)={X1(T), X2(T), ..., XM(T)}. The following three processes are divided according to different search results: S311. If the number of adjacent operating condition data sequences U of the current operating condition data sequence X(T) found in the training set constructed in step S21 is greater than 2, the adjacent operating condition data sequence X(w') closest to the current operating condition data sequence X(T) is selected, and the power output P(w') corresponding to the operating condition data sequence X(w') is used as the predicted value P*(T) of the power output of the current power generation equipment; S312. If no operating condition data sequence adjacent to X(T) is found in the training set constructed in step S21, at this time, in the historical monitoring data sequence, trace back a working condition data sequence of length L from the current T-th working condition data: XL(T) = {X(T-L), X(T-L-2), …, X(T-2), X(T-1)}, and its corresponding power output data sequence PL(T) = {P(T-L), P(T-L-2), …, P(T-2), P(T-1)}. In the historical monitoring data sequence, search for a working condition data sequence XL(t) similar to the current working condition data sequence XL(T), where XL(t) = {X(t-L), X(t-L-2), …, X(t-2), X(t-1)}. The working condition data sequences XL(t) and XL(T) are similar, and the conditions to be satisfied are: any element X(t-i) in XL(t) and the corresponding element X(T-i) in XL(T) are adjacent. The method for adjacent judgment uses the method for finding adjacent working condition data sequences in step S21. Denote the first similar working condition data sequence found as XL(t*), and XL(t*) is the working condition data sequence closest to the current time point T. Use the power output P(t*) corresponding to the working condition data sequence X(t*) as the predicted value P*(T) of the current power generation equipment power output. S313. No working condition data sequence similar to the current working condition data sequence XL(T) is found in the historical monitoring data sequence. At this time, the predicted value can be output using the power output machine learning model constructed in step S2. The specific operation is as follows: Input the current working condition data sequence X(T) = {X1(T), X2(T), …, XM(T)} into the trained multi-layer neural network. After the calculation of the neural network, obtain the predicted value P*(T) of the current power generation equipment power output. S4. According to the predicted value of the current power generation equipment power output output in step S3, abnormal detection and alarm can be performed. The specific operation is as follows: According to the predicted value P*(T) output by the prediction module and the monitored actual power output value P(T), judge whether the current monitored value is abnormal by whether it is a near neighbor. If it satisfies |P*(T) – P(T)| < STDP / R, then it is a near neighbor, and it is judged that no abnormality has occurred. Otherwise, it is judged that there is a potential hidden danger of abnormality, generate an alarm message text message, and send it to the operation and maintenance management personnel, requesting the operation and maintenance management personnel to conduct an inspection of the power generation equipment to eliminate the potential fault.

2. A clean energy power generation prediction method according to claim 1, characterized in that: In step S311, the distance between the current working condition data sequence X(T) and the adjacent working condition data sequence X(w) is calculated using the Euclidean distance calculation formula. Among them, the distance calculation formula between the working condition data sequence X(T) and the adjacent working condition data sequence X(w') is as follows: Deuclidean(X(T),X(w')) = min{Deuclidean(X(T),X(w)), and the working condition data X(w) adjacent to X(T)}.

3. A clean energy power generation prediction system, characterized in that: A clean energy power generation power prediction method as described in any one of claims 1-2 is adopted, comprising: an electric energy meter and sensor module, a historical monitoring database, a high-quality training set construction module, a power output neural network training module, a common operating condition feature prediction module, a high-dispersion operating condition feature prediction module, a special operating condition feature prediction module and a clean energy power generation anomaly detection module; wherein, The electric energy meter and sensor module are used to monitor and collect power output data and operating condition characteristic data of the power generation equipment in real time, and send the monitoring data to the historical monitoring database; The historical monitoring database is used to store the power output data and operating condition characteristic data sent by the electric energy meter and the sensor module; The high-quality training set construction module is used to extract high-quality training samples from the historical monitoring data sequence stored in the historical monitoring database based on the aggregation degree of the data, eliminate the interference of abnormal data samples, and train the power output neural network model; The power output neural network training module is used to train a multi-layer power output neural network based on a high-quality training set using a gradient descent neural network training algorithm, and fit the nonlinear relationship between power output and operating condition characteristics; The common operating condition feature prediction module is used to search for adjacent operating condition data sequences from a high-quality training set using the nearest neighbor prediction criterion and output a predicted value using the statistical average output power; The high-divergence operating condition characteristic prediction module is used to adopt the criterion of approximate data sequence prediction, calculate the similar operating condition data sequence matching the current operating condition data sequence from the historical monitoring data sequence, and use the power output of the most recent similar operating condition data sequence as the power output prediction value; The special operating condition characteristic prediction module is used to adopt a neural network model that fits the nonlinear relationship between power output and operating condition characteristics, and output a predicted value through neural network forward propagation; The clean energy power generation abnormality detection module determines whether the current detection value is abnormal based on the predicted value output by the prediction module and the actual power output value monitored. If it is abnormal, an alarm information SMS is generated and sent to the operation and maintenance management personnel.

4. An electronic device, characterized in that: The method comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 2 is completed.

5. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the method described in any one of claims 1-2.