A method for predicting power generation load

Abstract

The present invention provides a method for predicting power generation load, which obtains load prediction data within a preset time period from a database; normalizes the load prediction data; constructs a random forest model, calculates the relevant weights between the corresponding measurement points of the load prediction data, and uses them as initial weights; starts to construct an attention gate, and if the attention gate is constructed for the first time, inputs the initial weight into the attention gate, otherwise inputs the hyperparameters obtained by the flower pollination optimization algorithm into the attention gate; adds the constructed attention gate weight data to a GRU model for training; sets a threshold, calculates the error between the predicted value and the actual value; and obtains the best hyperparameters according to the flower pollination optimization algorithm. The present invention introduces an attention mechanism, which can not only handle long-term dependencies in historical time during prediction, but also can sample based on importance, thereby improving the accuracy of prediction. Sampling is performed based on the importance of historical data, and overly complex data preprocessing is not required, thereby improving operating efficiency.

Description

Technical Field

[0001] The present invention relates to the field of power generation load prediction, and in particular to a power generation load prediction method. Background Art

[0002] The grid-connected generator sets achieve a balanced relationship between power supply and demand based on the actual situation of power production and user consumption. Intelligent power generation load forecasting algorithms based on data mining are gradually being valued by researchers. As an important task in the current construction of smart grids, the accuracy of the prediction algorithm will have a huge impact on economic dispatch, real-time control, operation plans and development plans. It is also one of the important criteria for measuring whether the management level of power companies has reached modernization.

[0003] In the problem of power generation load forecasting, classic algorithms include support vector machine forecasting, neural network forecasting, autoregressive moving average model (ARIMA) and other methods. For example, in the disclosed invention patent "A Power Load Forecasting Method", patent application publication number CN 109934375 A, the power load data is clustered and analyzed by the DBSCAN clustering algorithm, and different types of power load data are automatically divided, and then used as training data to build an LSTM load forecasting model. Real-time similar category data and original data are merged as inputs of the LSTM model for the prediction of daily power load curves. This load forecasting method takes into account the use of deep learning models to learn and use similar pattern data information, but does not consider the strength of the correlation between different independent variables and the target variable.

[0004] In the prior art, there is also a method of analyzing the similarity between current load data and historical load data, and using the load change trend of the most similar day to predict future loads. This load forecasting method mainly uses the similarity of pattern data for prediction. If there is a load data pattern that does not exist in history, the prediction method will have a false alarm. Only the most similar load data pattern is used as the prediction result, and other pattern data that meets the similarity are not extracted. The prediction result lacks certain rationality. Summary of the invention

[0005] The present invention provides a power generation load forecasting method. The forecasting method focuses on how to fully explore the forecasting relationship between power generation load and related variables and effectively measure the forecasting accuracy. It attempts to better solve the problem of low accuracy of power plant load forecasting by establishing a new power generation load forecasting technology, thereby facilitating better operation and management of auxiliary power plants.

[0006] A method for predicting power generation load, the method comprising:

[0007] Step 1: Obtain load forecast data within a preset time period from the database;

[0008] Step 2: Normalize the load forecast data;

[0009] Step 3: Construct a random forest model based on Gini coefficient calculation and calculate the relevant weights between the corresponding measurement points of the load forecast data as the initial weights;

[0010] Step 4: Start building the attention gate. If it is the first time to build the attention gate, input the initial weight into the attention gate. Otherwise, input the hyperparameters obtained by the flower pollination optimization algorithm into the attention gate.

[0011] Step 5: Establish a GRU model and add the attention gate weight data constructed in step 4 to the GRU model for training;

[0012] Step 6: Set the threshold T thre , calculate the error between the predicted value and the actual value, if the error F error If it is greater than the threshold, go to step 7, otherwise output the optimal GRU model;

[0013] Step 7: Obtain the best hyperparameters based on the flower pollination optimization algorithm.

[0014] The method of the present invention further comprises:

[0015] Step 11: Obtain weather forecast data for a preset time period in the future from the database as independent variable data;

[0016] Step 12: Normalize the independent variable data according to the training normalization model;

[0017] Step 13: Load the optimal GRU model output from step 6 of the training phase;

[0018] Load the optimal GRU model and obtain the optimal network model parameters;

[0019] Step 14: Input the preprocessed independent variable data into the optimal GRU model and run the model;

[0020] Get the optimal network model parameters, load the data, and calculate the predicted time by formulas (20), (21), and (22) to reset the gate and update the gate status;

[0021]

[0022] r t =σ(w r x t +u r h t-1 ) (twenty one)

[0023] z t =σ(w z x t+u z h t-1 ) (twenty two)

[0024] Finally, the hidden layer state combination is obtained;

[0025] Step 15: Calculate the load forecast value F of the output independent variable data forecast ;

[0026] The input data is calculated according to formula (23) for the hidden layer h at the current moment t Later, we can get the output of the GRU network model, w y is the weight;

[0027] F forecast =σ(w y h t ) (1).

[0028] It can be seen from the above technical solutions that the present invention has the following advantages:

[0029] The power generation load prediction method provided by the present invention is based on clustering of historical load data to predict the load data of the unit for a preset time in the future, and can predict the load of a preset time in the future according to demand.

[0030] The present invention introduces an attention mechanism, which can not only handle long-term dependencies in historical time during prediction, but also can sample based on importance, thereby improving the accuracy of prediction.

[0031] The present invention performs sampling based on the importance of historical data, does not require overly complex data preprocessing, and improves operating efficiency.

[0032] In terms of historical data weights, the present invention first uses the normalized random forest node splitting impurity change, and then performs optimization adjustment through the flower pollination optimization algorithm, so that the load forecasting calculation process can reflect the real-time mutual influence changes between parameters.

[0033] The GRU model involved in the present invention is a variant of the RNN model, which has the advantages of fewer parameters, fast convergence speed, and fast iteration. The present invention can improve the operating efficiency.

[0034] The present invention is based on the periodicity of short-term power loads. At the same time, the existing power grid power generation is greatly disturbed by external factors. The meteorological data precipitation, air pressure, wind speed, temperature, humidity, comprehensively reflect the influence of hydropower, wind power, and photovoltaic power, and more carefully reverse the current trend of networked power generation. At the same time, the power consumption is closely related to the temperature, and it is also very important to use meteorological factors as an important indicator for load prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the description 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.

[0036] Figure 1 It is a flow chart of the power generation load forecasting method;

[0037] Figure 2 It is the operation flow chart of the present invention;

[0038] Figure 3 This is the prediction effect diagram;

[0039] Figure 4 Schematic diagram of pollinating flowers. DETAILED DESCRIPTION

[0040] 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.

[0041] In the power generation load forecasting method provided by the present invention, the units and algorithm steps of each example described in the disclosed embodiment can be implemented by electronic hardware, computer software or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0042] In the power generation load forecasting method provided by the present invention, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, or it can be an electrical, mechanical or other form of connection.

[0043] In the power generation load forecasting method provided by the present invention, the described features, structures or characteristics can be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to give a full understanding of the embodiments of the present invention. However, those skilled in the art will appreciate that the technical solution of the present invention can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. can be adopted. In other cases, well-known methods, devices, implementations or operations are not shown or described in detail to avoid obscuring various aspects of the present invention.

[0044] The power generation load forecasting method based on the FPA-GRU algorithm provided by the present invention is a new load forecasting technical method based on historical load data and meteorological data. The present invention combines the flower pollination optimization algorithm and the GRU algorithm with an attention mechanism based on historical load data to realize the load forecasting function under the optimal parameter weight. The method of the present invention mainly includes two processes: establishing a model and running the model.

[0045] The FPA-GRU load forecasting model training process involved in the present invention includes:

[0046] Figure 1 The flow chart of the model established for the present invention, the whole modeling process mainly includes the following steps:

[0047] Step 1: Obtain all historical data of the specified time period of the specified parameters of the load forecast from the database, identify and eliminate the shutdown data in all the data according to the start and stop, and use the processed data as historical data. The specified parameters of the load forecast can be the load forecast data within the preset time period described in the present invention, and of course the load forecast data within the preset time period here is the historical data.

[0048] The present invention takes the No. 1 generating unit of a power plant as an example, and the operation process is as follows: first, select the parameters associated with the unit load: precipitation, air pressure, wind speed, temperature, humidity, historical load of Unit 1 and other measurement points in the area where the power plant is located, and read the historical data from 2019-7-14 10:00 to 2020-11-24 10:00 from the real-time database of the power plant based on the above information.

[0049] A set of data samples with 6 measurement points and m time points can be regarded as an n-dimensional vector at time j, expressed as:

[0050] u(t m )=[u j1 ,u j2 ,u j3 ,…,l j6 ]

[0051] Among them, u j1 ,uj2 ,u j3 ,…,l j6 It represents precipitation, air pressure, wind speed, temperature, humidity, and historical load respectively.

[0052] The data file is generally in the form of an m*n matrix, and the specific form is as follows

[0053]

[0054] Step 2: Normalize the historical data to eliminate the impact of data dimension;

[0055] For the data F at time a a Perform data normalization to obtain the normalized data matrix F a-norm According to the following formula, the training data F a Map data to the [0,1] interval

[0056]

[0057] Among them, F a-norm is the normalized value at time a, F a-max is the maximum value of this parameter, F a-min is the minimum value of this parameter.

[0058] Step 3: Construct a random forest model based on the Gini coefficient calculation, and calculate the correlation weights between each measurement point of the historical data as the initial weights;

[0059] The Gini coefficient refers to the probability of randomly selecting a sub-item from a data set and measuring its probability of being incorrectly divided into other parts.

[0060] 3.1 Calculate the Gini index score of each test point of each decision tree, including the node split impurity.

[0061]

[0062] Among them, m represents the mth measurement point, K represents that the m measurement points contain K categories, and Pmk represents the proportion of category k in node m.

[0063] Then calculate the weight of the feature in the node:

[0064]

[0065] Among them, GIi and GIr represent the Gini index of the two new nodes after branching.

[0066] 3.2 Accumulate and count the Gini index scores of each measuring point to find the average score, normalize it, and use the normalized score as the initial weight between the various measuring points in the historical data.

[0067] If feature X j In the node set M that appears in decision tree i, then X j The reduction in impurity in the i-th tree is:

[0068]

[0069] Where i represents the i-th decision tree, j represents the j-th measurement point of the i-th tree, and M represents the set of measurement points of the decision tree;

[0070] Assuming that the random forest has a total of n trees, the reduced impurity of each test point in the random forest is:

[0071]

[0072] Among them, j is the jth node, and n is the total number of decision trees;

[0073] Finally, all the obtained importance scores are normalized to obtain the weights:

[0074]

[0075] Step 4: Start building the attention gate. If it is the first time, input the initial weight into the attention gate. Otherwise, input the hyperparameters obtained by the flower pollination optimization algorithm into the attention gate.

[0076] The GRU model helps capture long-term dependencies, but it is not enough to pay different attention to sub-window features in multiple time steps. To address this problem, a layer of "attention gate" is artificially constructed outside the GRU model, similar to the learning method of evolutionary attention. The importance-based sampling mode is confirmed in the mining of temporal relationships.

[0077] According to the attention gate, the input data is sampled for importance, and the attention gate weight data is:

[0078] w=[w1, w2, w3,..., w6].

[0079] Step 5: Establish a GRU model and add the attention gate weight data constructed in step 4 to the GRU model for training;

[0080] GRU has the ability to model long-term historical information of time series, and GRU network is selected to build load forecasting network. GRU has reset gate and update gate.

[0081] Using the attention gate weights in step 4, the input data is adopted based on importance and sent to the GRU network. The nonlinear mapping function of the GRU unit calculation process can be learned through the following formula:

[0082] The GRU neural network mainly uses a GRU composed of two sigmoid functions and multiplication operations to control the selection of information, and uses the sigmoid function to transform the memory loading and the current input data to between 0 and 1, thereby realizing the memory and discarding of data.

[0083] In the GRU neural network model, the hidden layer state at the current moment is the sum of the hidden layer state at the previous moment and the current hidden layer activation state, that is:

[0084]

[0085] The reset gate and update gate are the combination of the current moment input data and the previous moment hidden layer state:

[0086] r t =σ(w r x t +u r h t-1 ) (8)

[0087] z t =σ(w z x t +u z h t-1 ) (9)

[0088] Where: σ is the sigmoid function, which can be used to t and z t Limited to 0-1, w r 、u r 、w z 、u z is the neural network weight.

[0089] Reset Gate t The value h of the hidden layer at the previous moment can be t-1 Selectively reset, that is, the reset gate determines the degree of retention in the hidden layer at the previous moment, r t The closer the value is to 1, the more data is retained in the hidden layer at the previous moment. The hidden layer data at the previous moment is processed by the reset gate and is consistent with the current input x. t Combined with the Tanh activation function, the activation state of the hidden layer at the current moment can be obtained

[0090]

[0091] n z =wx t +u(r t ⊙h t-1 ) (11)

[0092] Where: ⊙ represents the product operation, W and u are the weights of the GRU neural network.

[0093] h t The update process is mainly composed of the update gate z t Done, z t The closer the value is to 0, the greater the degree of information retention at the current moment.

[0094] The optimized attention weights continuously improve the performance without falling into local optimality. The attention distribution of the input driving sequence in multiple real time steps of each data also shows different scales, which effectively utilizes the local information within the sampling window.

[0095] Step 6: Set the threshold T thre , calculate the predicted value F forecast With the actual value F actual If the error F error If it is greater than the threshold, go to step 7, otherwise output the optimal GRU model;

[0096] F error =F forecast -F actual (12)

[0097] Step 7: Obtain the best hyperparameters based on the flower pollination optimization algorithm. Figure 3 and Figure 4 As shown, return to step 4;

[0098] 7.1 Initialization parameters, flower population number n, transition probability p;

[0099] 7.2 Calculate the fitness value of each solution and solve the current optimal solution and optimal value;

[0100] 7.3 If the condition of transition probability p>rand is met, the solution is updated according to the following formula and out-of-bounds processing is performed.

[0101]

[0102] in: are the solutions of the t+1th and tth generations respectively; θ represents the step length control amount, which is generally 1. represents point-to-point multiplication, Levy is the Levy flight search path, β ranges from [1, 3], and is generally 3 / 2. The calculation formula of Levy is:

[0103]

[0104] Where: λ ranges from [0, 2], usually 3 / 2, Γ(λ) is the standard gamma function, s is the search path, and the calculation formula of S is:

[0105]

[0106]

[0107]

[0108] Where μ, v are random numbers that follow a normal distribution

[0109] 7.4 If the condition p>rand does not hold, update the optimal solution according to the following formula and perform out-of-bounds processing:

[0110]

[0111] Where: ∈ is a random number that follows a uniform distribution on [0, 1], It's the same plant.

[0112] 7.5 Calculate the fitness value corresponding to the new solution obtained in 7.3 or 7.4. If the fitness of the new solution is good, replace the current solution and the current fitness value with the new solution and the fitness value corresponding to the new solution respectively; otherwise, retain the current solution and the current fitness value.

[0113] 7.6 If the fitness value corresponding to the new solution is better than the global optimal value, update the global optimal solution and the global optimal value.

[0114] 7.7 Determine the end condition. If it is met, output the optimal solution, otherwise return to 7.3.

[0115] Figure 2 As shown, it is a flow chart of the operation of the model of the present invention, and the whole process mainly includes the following steps:

[0116] Step 1: Obtain the weather forecast data for the next 24 hours from the database as the independent variable data.

[0117] Get the weather data for the next 24 hours at the forecast time F b : Precipitation, air pressure, wind speed, temperature, and humidity are used as independent variable data;

[0118] The five columns are precipitation, air pressure, wind speed, temperature, and humidity.

[0119]

[0120] Step 2: Normalize the independent variable data according to the training normalization model;

[0121] For data F b Perform data normalization to obtain the normalized data matrix F b-norm According to the following formula, the training data F b Map the data to the interval [0, 1].

[0122]

[0123] Among them, F b-norm is the normalized value of time, F b-max is the maximum value of the corresponding measurement point of the normalized model, F b-min is the minimum value of the corresponding measurement point of the normalized model.

[0124] Step 3: Load the optimal GRU model output from step 6 of the training phase;

[0125] Load the optimal GRU model and obtain the optimal network model parameters.

[0126] Step 4: Input the preprocessed independent variable data into the optimal GRU model and run the model;

[0127] Get the optimal network model parameters, load the data, and calculate the predicted time by formulas (20), (21), and (22) to reset the gate and update the gate state

[0128]

[0129] r t =σ(w r x t +u r h t-1 ) (twenty one)

[0130] x t =σ(w z x t +u z h t-1 ) (twenty two)

[0131] Finally, the hidden layer state combination is obtained;

[0132] Step 5: Calculate the load forecast value F of the output independent variable data forecast ;

[0133] The input data is calculated according to formula (23) for the hidden layer h at the current moment t Later, we can get the output of the GRU network model, w y is the weight.

[0134] F forecast =σ(w y h t ) (twenty three)

[0135] As an embodiment of the present invention, the online prediction process of the #1 generator set is as follows:

[0136] Obtain the data for 24 hours after 10:00 on November 24, 2020 from the database, select time, precipitation, air pressure, wind speed, temperature, and humidity as features, perform data preprocessing and load the model. Input the data into the model, run the model, and output load forecast data F forecast =[u1,u2,…u n ].

[0137] Compared with the traditional method, this method has higher operating efficiency, and the maximum prediction error of the unit load is less than 7%. It can guide the optimization of the start and stop of important auxiliary equipment of the power plant generator set, which helps to improve the economic operation level of the unit. Short-term load forecasting is an important means of operating and controlling the generator set. The present invention is of great significance to the scheduling of power generation capacity, reliability analysis, maintenance of generator set equipment and other operational decisions.

[0138] The power generation load forecasting method based on the FPA-GRU algorithm provided by the present invention is a combination of the units and algorithm steps of each example described in the embodiments disclosed herein, and can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0139] The power generation load forecasting method based on the FPA-GRU algorithm provided by the present invention can be written in any combination of one or more programming languages ​​to write program codes for performing the operations disclosed herein, and the programming languages ​​include object-oriented programming languages-such as Java, C++, etc., and also include conventional procedural programming languages-such as "C" language or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, as an independent software package, partially on the user computing device and partially on the remote computing device, or completely on the remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (for example, using an Internet service provider to connect through the Internet).

[0140] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting power generation load, characterized in that: Methods include: Step 1: Obtain load forecast data within a preset time period from the database; Step 2: Normalize the load forecast data; Step 3: Construct a random forest model based on Gini coefficient calculation and calculate the relevant weights between the corresponding measurement points of the load forecast data as the initial weights; Step 4: Start building the attention gate. If it is the first time to build the attention gate, input the initial weight into the attention gate. Otherwise, input the hyperparameters obtained by the flower pollination optimization algorithm into the attention gate. Step 5: Establish a GRU model and add the attention gate weight data obtained in step 4 to the GRU model for training; Step 6: Set the threshold , calculate the load forecast value With the actual value F actual If the error If it is greater than the threshold, go to step 7, otherwise output the optimal GRU model; Step 7: Obtain the best hyperparameters based on the flower pollination optimization algorithm; In step 4, According to the attention gate, the load forecast data is sampled for importance and the attention gate weight data is obtained: ； The method also includes: Step 11: Obtain weather forecast data for a preset time period in the future from the database as independent variable data; Step 12: Normalize the independent variable data according to the training normalization model; Step 13: Load the optimal GRU model output from step 6 of the training phase; Load the optimal GRU model and obtain the optimal network model parameters; Step 14: Input the preprocessed independent variable data into the optimal GRU model and run the model; Get the optimal network model parameters, load the data, and calculate the predicted time by formulas (20), (21), and (22) to reset the gate and update the gate status; Finally, the hidden layer state combination is obtained; Step 15: Calculate the load forecast value of the output independent variable data ; The input data is calculated according to formula (23) for the hidden layer at the current moment. Later, the output of the GRU network model can be obtained. is the weight; ； In step 3, Calculate the gini index score of each measurement point of each decision tree; Among them, m represents the mth measurement point, K represents that the m measurement points contain K categories, and P mk Indicates the proportion of category k in measurement point m; Calculate the weight of the feature in the node: Among them, GI I and GI r Respectively represent the Gini index of the two new nodes after branching; The Gini index scores of each measuring point are accumulated and averaged, and then normalized. The normalized scores are used as the initial weights between the various measuring points in the historical data. If the feature In the node set M that appears in decision tree i, then The reduction in impurity in the i-th tree is: Among them, i represents the i-th decision tree, j represents the j-th measurement point of the i-th tree, and M represents the set of measurement points of the decision tree; assuming that the random forest has a total of n trees, the reduced impurity of each measurement point in the random forest is: Among them, j is the jth measurement point, and n is the total number of decision trees; Finally, all the obtained importance scores are normalized to obtain the weights: In step 5, Using the attention gate weights from step 4, the load forecast data is fed into the GRU neural network; The GRU neural network has two Function and multiplication operation synthesized GRU to control the selection of information, using The function converts the memory load and the current input data to between 0 and 1; In the GRU neural network model, the hidden layer state at the current moment is the sum of the hidden layer state at the previous moment and the current hidden layer activation state, that is: (7) The reset gate and update gate are the combination of the current moment input data and the previous moment hidden layer state: （8） （9） in: yes function, and Limited to between 0-1, , , , is the neural network weight; Reset Gate The value of the hidden layer at the previous moment Reset, reset the gate configuration to the extent retained in the hidden layer at the previous moment, The closer the value is to 1, the more data is retained in the hidden layer at the previous moment. The hidden layer data at the previous moment is processed by the reset gate and the current input Combined and through Activation function, get the activation state of the hidden layer at the current moment ; （10） （11） Where: represents the product operation, , is the weight of the GRU neural network; The update process is performed by the update gate Finish, The closer the value is to 0, the greater the degree of information retention at the current moment; In step 7, 7.1 Initialization parameters, flower population number n, transition probability p; 7.2 Calculate the fitness value of each solution and find the current optimal solution and optimal value; 7.3 If the condition of transition probability p > rand is met, update the solution according to the following formula and perform out-of-bounds processing; （13） in: , are the solutions of the t+1th generation and the tth generation respectively; Represents the step length control value, the value is 1, represents point-to-point multiplication, is the Levy flight search path, Take 3 / 2; Levy's calculation formula: (14) in: The value is 3 / 2, is the standard gamma function, s is the search path, and the calculation formula of s is: (15) (16) (17) in , is a random number that follows a normal distribution; 7.4 If the condition p > rand does not hold, update the optimal solution according to the following formula and perform out-of-bounds processing: in: is a random number that follows a uniform distribution on [0,1]. , are the same value; 7.5 Calculate the fitness value corresponding to the new solution obtained in 7.3 or 7.

4. If the fitness of the new solution is good, replace the current solution and the current fitness value with the new solution and the fitness value corresponding to the new solution respectively. Otherwise, keep the current solution and the current fitness value. 7.6 If the fitness value corresponding to the new solution is better than the global optimal value, update the global optimal solution and the global optimal value; 7.7 Determine the end condition. If it is met, output the optimal solution, otherwise return to 7.

3.

2. The power generation load prediction method according to claim 1, characterized in that: In step 1, the shutdown data in the load forecast data within a preset time period is identified and eliminated based on the start and stop, and the load forecast data after elimination is used as historical data in steps 2 to 7.

3. The power generation load prediction method according to claim 1, characterized in that: In step 2, Load forecast data for time a Perform data normalization to obtain a normalized data matrix , according to the following formula for the training data Map data to Interval (1) in, is the normalized value at time a, is the maximum value, is the minimum value.

4. The power generation load prediction method according to claim 1, characterized in that: The method also includes: Step 11 also includes: Get weather data for a preset time period in the future : Precipitation, air pressure, wind speed, temperature, and humidity are used as independent variable data; The five columns are precipitation, air pressure, wind speed, temperature, and humidity. 。 5. The power generation load forecasting method according to claim 1, characterized in that: The method also includes: In step 12, the data Perform data normalization to obtain a normalized data matrix ; According to the following formula, the training data Map data to interval; (19) in, is the normalized value of time, is the maximum value of the corresponding measurement point of the normalized model, is the minimum value of the corresponding measurement point of the normalized model.

Images