Source load power prediction method based on prediction error correction, medium and equipment

Abstract

According to the source load power prediction method based on prediction error correction, the medium and the equipment, the historical monitoring data of the equipment is acquired, the training set, the test set and the prediction set are constructed, and the neural network model taking the historical monitoring data as input and the source load power prediction data as output is constructed and trained according to the training set; and according to the test set and the trained neural network model, determining an error between prediction data and real data in the test set to obtain an error probability distribution curve, carrying out random sampling on the error probability distribution curve to obtain prediction error compensation data, and according to the prediction set and the prediction error compensation data, obtaining final source load power prediction data. The problems that in the prior art, source load uncertainty cannot be effectively handled, and prediction precision is insufficient are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power control, and in particular to a source-load power prediction method, medium and device based on prediction error correction. Background Art

[0002] With the intensification of the global energy crisis and the increasing prominence of environmental pollution, the traditional energy system faces huge challenges. New renewable energy sources such as solar energy and wind energy have gradually become an important direction of global energy transformation due to their clean and renewable characteristics. In my country, new energy sources with solar energy and wind energy as the main body have developed rapidly and have become the absolute main body of new power installed capacity, and have shown a good development trend of scale and marketization across the country. However, with the increasing diversification of distributed power sources and load types, there are significant differences in the power generation characteristics and power generation costs of micro-sources in different energy forms, especially the power generation characteristics and load curves of renewable energy sources such as wind energy and solar energy are greatly affected by external objective factors such as meteorology, showing strong uncertainty.

[0003] In the current field of source-load (power source and load) power forecasting, most of the existing forecasting methods fail to fully consider the impact of source-load uncertainty on the forecast results, and also lack in-depth analysis of the causes of forecast errors and effective countermeasures. This makes it difficult for existing methods to accurately predict the changing trend of source-load power, especially in the context of increasing penetration of new energy, which brings many new challenges to grid system operators. Accurate source-load power forecasting is crucial to ensuring the safe, stable and economical operation of the power grid. Therefore, there is an urgent need for a new forecasting method that can effectively deal with source-load uncertainty and improve forecasting accuracy.

[0004] Therefore, how to improve it is a technical problem that needs to be solved urgently in this field. Summary of the invention

[0005] Based on this, the purpose of this application is to provide a source-load power prediction method, medium and device based on prediction error correction to solve at least one technical problem mentioned in the above background technology.

[0006] In a first aspect, the present application provides a source-load power prediction method based on prediction error correction, comprising:

[0007] Obtain historical monitoring data of the equipment and build training sets, test sets, and prediction sets;

[0008] Construct and train a neural network model based on the training set, taking historical monitoring data as input and source load power prediction data as output;

[0009] According to the test set and the trained neural network model, the error between the predicted data and the real data in the test set is determined, and the error probability distribution curve is obtained;

[0010] Perform random sampling on the error probability distribution curve to obtain prediction error compensation data;

[0011] Based on the prediction set and the prediction error compensation data, obtain the final source-load power prediction data.

[0012] Furthermore, the neural network model includes:

[0013] An input layer for receiving device historical monitoring data;

[0014] A feature extraction layer connected to the input layer, which is used to generate n input vector sequences according to the historical monitoring data and extract the time series features of each input vector sequence; n is the number of time nodes of the historical monitoring data;

[0015] A feature fusion layer connected to the feature extraction layer, which is used to configure weights for each time series feature to generate weighted vectors;

[0016] An output layer connected to the feature fusion layer, which is used to output source-load power prediction data according to the weighted vectors.

[0017] Furthermore, the feature extraction layer includes:

[0018] A generation unit for generating n input vector sequences according to the historical monitoring data;

[0019] And n GRU units connected in chronological order; the GRU unit at time t has two input vectors: the input vector sequence at the current moment and the output of the previous GRU unit; it is used to generate a hidden state sequence as the time series feature of the current input vector according to the input vector sequence.

[0020] Furthermore, the generation unit includes:

[0021] A division subunit connected to the input layer, which divides the historical monitoring data into input vector sequences at m moments according to the time node interval; m = n / 2;

[0022] An analysis subunit connected to the output of the division subunit, which is used to analyze the day-night data of each input vector sequence, divide each input vector sequence into a day input vector sequence and a night input vector sequence, obtain n input vector sequences, and input them into n GRU layers.

[0023] Furthermore, the feature fusion layer is an attention layer connected to the GRU layer, which is used to perform weighted processing on the hidden state sequence output by the GRU layer to obtain attention weights, so as to generate weighted vectors according to the hidden state sequence and the attention weights.

[0024] Furthermore, the steps of obtaining the error probability distribution curve include:

[0025] Input the test set into the neural network model to obtain the predicted source-load power data of the corresponding device within a set time period;

[0026] Collect the actual source-load power data of the device within a set time period to obtain the prediction error data between the actual data and the predicted source-load power data;

[0027] Resample the prediction error data to obtain a data sample set;

[0028] Perform kernel density estimation on the data sample set to obtain the error probability distribution curve.

[0029] Further, the step of resampling the prediction error data to obtain a data sample set includes:

[0030] Resample the prediction error data to obtain several data groups;

[0031] Arrange the elements in each data group in a set order to obtain several updated data groups;

[0032] Obtain the arithmetic mean of the elements in each column of the updated data groups to obtain the data sample set.

[0033] Further, the step of randomly sampling the error probability distribution curve to obtain the prediction error compensation data includes:

[0034] Randomly sample the error probability distribution curve to obtain several error samples;

[0035] Denormalize each error sample to obtain several actual error compensation amounts;

[0036] Construct an error compensation data set based on the actual error compensation amounts to obtain the prediction error compensation data.

[0037] In a second aspect, the present application also provides a computer storage medium storing executable program code; the executable program code is used to execute the source-load power prediction method based on prediction error correction according to any one of the first aspect.

[0038] In a third aspect, the present application also provides a terminal device including a memory and a processor; the memory stores program code executable by the processor; the program code is used to execute the source-load power prediction method based on prediction error correction according to any one of the first aspect.

[0039] A source-load power prediction method based on prediction error correction of the present invention obtains historical monitoring data of devices, constructs a training set, a test set, and a prediction set to provide a basis for subsequent steps. Then, a neural network model with historical monitoring data as input and source-load power prediction data as output is constructed and trained based on the training set, so as to obtain the temporal features and important information in the prediction set through the neural network model and obtain initial prediction data. Then, according to the test set and the trained neural network model, the error between the prediction data and the real data in the test set is determined to obtain an error probability distribution curve, and then random sampling is performed on the error probability distribution curve to obtain prediction error compensation data, providing data support for subsequent error correction by analyzing the prediction error and estimating the probability distribution, improving the accuracy of the error distribution. Finally, according to the prediction set and the prediction error compensation data, the final source-load power prediction data is obtained, and the initial prediction data is directly corrected by the prediction error compensation data, so as to effectively cope with the uncertainty of source-load and improve the prediction accuracy. It solves the problems that the prior art cannot effectively cope with the uncertainty of source-load and has insufficient prediction accuracy, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a framework diagram of the source-load power prediction method based on prediction error correction according to an embodiment of the present invention;

[0041] Figure 2 It is a flowchart of the source-load power prediction method based on prediction error correction according to an embodiment of the present invention;

[0042] Figure 3 It is a schematic diagram of the neural network model structure according to an embodiment of the present invention;

[0043] Figure 4 It is a comparison diagram of the power generation prediction data of a wind power generation device according to an embodiment of the present invention;

[0044] Figure 5 It is a comparison diagram of the power generation prediction data of a wind power generation device according to an embodiment of the present invention;

[0045] Figure 6 It is a comparison diagram of the load power prediction data according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0047] It should be noted that if there are directional indications involved in the embodiments of the present invention, such as up, down, left, right, front, back..., then the directional indications are only used to explain the relative positional relationship, movement conditions, etc. between components in a specific posture. If the specific posture changes, the directional indications will also change accordingly. Additionally, if there are descriptions such as "first, second", "S1, S2", "step one, step two", etc. involved in the embodiments of the present invention, such descriptions are only for descriptive purposes and should not be construed as indicating or implying their relative importance or implicitly indicating the quantity of the indicated technical features or indicating the execution order of the method, etc. Those skilled in the art can understand that all those that do not violate the inventive concept of the invention should be included in the protection scope of the present invention.

[0048] With the diversification of distributed power sources and load types, the power generation characteristics and costs of different energy forms of micro-sources are not the same, and the power generation of renewable energy sources such as wind and light and the load curve are greatly affected by external objective factors such as meteorology and show great uncertainty. The current source-load power prediction methods do not fully consider the impact of source-load uncertainty on prediction, nor do they analyze the causes of prediction errors and corresponding countermeasures, making it difficult to accurately predict the change trend of source-load power. Therefore, based on the power prediction method framework as Figure 1 shown, the present invention provides a source-load power prediction method based on prediction error correction as Figure 2 shown to reduce the error of power prediction:

[0049] S1: Obtain the historical monitoring data of the device, and construct a training set, a test set, and a prediction set;

[0050] Specifically, optionally but not limited to, according to the device for which source-load (power source and load) power prediction is required, obtain the historical monitoring data of the current device, and construct a training set, a test set, and a prediction set based on the obtained historical monitoring data of the device. Preferably, perform normalization processing on the training set, the test set, and the prediction set to accelerate the convergence speed and improve the training speed of the neural network model.

[0051] Exemplarily, taking the source-load (power source and load) power prediction of a wind turbine as an example, optionally obtain the historical hourly power generation of a single wind turbine, the historical hourly power generation of a single photovoltaic panel, the historical hourly power consumption of regional users, etc. as basic data, and divide the basic data into a training set, a test set, and a planned power prediction set; among them, the training set is used to train the neural network model, and the test set is used to obtain the error between the predicted value and the true value of the neural network model.

[0052] S2: Construct and train a neural network model with historical monitoring data as the input and source-load power prediction data as the output according to the training set;

[0053] Specifically, the neural network model can adopt any neural network structure in the prior art to construct an initial neural network model; then, the training set is input into the initial neural network model to optimize the parameters of the initial neural network model, and a final neural network model is obtained. Preferably, the initial neural network model adopts an Att-GRU neural network structure.

[0054] Further preferably, as Figure 3 shown, the specific structure of the neural network model may optionally but not limited to include:

[0055] An input layer for receiving historical monitoring data of the device;

[0056] A feature extraction layer, connected to the input layer, for generating n input vector sequences according to the historical monitoring data and extracting the temporal features of each input vector sequence; n is the number of time nodes of the historical monitoring data;

[0057] A feature fusion layer, connected to the feature extraction layer, for configuring weights for each temporal feature to generate weighted vectors;

[0058] An output layer, connected to the feature fusion layer, for outputting source-load power prediction data according to the weighted vectors.

[0059] Specifically, the feature extraction layer includes:

[0060] A generation unit for generating n input vector sequences according to the historical monitoring data;

[0061] And, n GRU units connected in chronological order; the GRU unit at time t has two input vectors: the input vector sequence x t at the current moment and the output h t-1 of the previous GRU unit; for generating a hidden state sequence {h t} according to the input vector sequence {X t} as the temporal feature of the current input vector.

[0062] More specifically, the generation unit includes:

[0063] A division sub-unit, connected to the input layer, divides the historical monitoring data into input vector sequences at m moments according to the time node interval; m = n / 2; preferably, the historical monitoring data is divided into input vector sequences at m moments in units of days.

[0064] An analysis sub-unit, connected to the output of the division sub-unit, for analyzing the day-night data of each input vector sequence, dividing each input vector sequence into a day input vector sequence and a night input vector sequence, obtaining n input vector sequences, and inputting them into n GRU layers, thereby obtaining a hidden state sequence {h t}{It is the time series feature of the current input vector. Since the wind power resources and light conditions vary greatly at different times of the day, for example, at 12 o'clock and 24 o'clock, the input vector sequences can be divided into day input vector sequences and night input vector sequences according to different times of each day, obtaining n input vector sequences. Subsequently, the special fusion layer can configure attention weights for different input vector sequences according to the characteristics of each data to improve the accuracy of subsequent predictions.

[0065] More specifically, the feature fusion layer is an attention layer, connected to the GRU layer, and is used to perform weighted processing on the hidden state sequence {h t}} to obtain the attention weight β t , so as to generate a weighted vector A t according to the hidden state sequence {h t} and the attention weight β t .

[0066] More specifically, the output layer is connected to the attention layer and is used to generate the initial prediction data PI of the source-load power through a fully connected network based on the weighted vector A t .

[0067] Exemplarily, after receiving the historical monitoring data of the device, the feature extraction layer divides the historical monitoring data of the device by time to obtain the input vector sequence {X t}(t = 1, 2,..., n); the GRU layer contains several GRU units, the GRU units contain a reset gate and an update gate, and each GRU unit is connected in chronological order. Each GRU unit receives the input vector X t at the current moment and the output hidden state h t-1 of the previous GRU unit, and then outputs the current hidden state h t . Thus, according to all the hidden states h t , the hidden state sequence {h t} is constructed.

[0068] Preferably, the internal calculation process of the GRU layer can be as shown in Equations 2-1, 2-2, 2-3, and 2-4:

[0069] r t = σ(W r * [h t-1 , X t + b r ) 2-1

[0070] z t = σ(W z * [h t-1 , X t + b z ) 2-2

[0071]

[0072] Among them, r t and Z t are the outputs of the reset gate and update gate of each GRU unit respectively. W r 、W z 、W h are weight matrices, b r 、b z 、b h are bias terms, σ is the Sigmoid function, φ is the Tanh function, ⊙ represents element-wise multiplication, h t-1 is the output hidden state of the previous GRU unit, is the candidate matrix, which contains all vectors that may be the output of the GRU layer, and h t is the output hidden state of the current GRU unit.

[0073] Preferably, the internal calculation process of the feature fusion layer, that is, the attention layer, can be optionally as shown in Equations 2-5, 2-6, and 2-7:

[0074] β t = qtanh(ωh t + b) 2-5

[0075]

[0076] Among them, β t represents the attention probability value corresponding to ht, q and ω are weight coefficients, b is the bias coefficient, At is the weighted vector, corresponding to the source load power prediction data at time t.

[0077] S3: According to the test set and the trained neural network model, determine the error between the prediction data and the real data in the test set to obtain the error probability distribution curve

[0078] Input the test set into the neural network model to obtain the source load power prediction data, and obtain the error between the prediction data and the real data of the corresponding device to obtain the error probability distribution curve;

[0079] Specifically, it is optional but not limited to input the input data of the test set into the trained neural network model to obtain the source load power prediction data of the corresponding device after a period of time, and then obtain the prediction error between the source load power prediction data and the real source load power data according to the real source load power data of the corresponding device during this period, so as to obtain the error probability distribution curve.

[0080] Exemplarily, assume that the test set contains historical monitoring data of the device at six o'clock, seven o'clock, eight o'clock, nine o'clock, and ten o'clock. Optionally, use the historical monitoring data of the device at six o'clock, seven o'clock, eight o'clock, and nine o'clock as the input data set, and input it into the trained neural network model to obtain the predicted source-load power data of the corresponding device at ten o'clock. Then, compare it with the historical monitoring data of the device at ten o'clock to obtain the prediction error between the predicted source-load power data and the actual source-load power data, so as to further obtain the error probability distribution curve. Preferably, the input data set can be arbitrarily set by those skilled in the art. That is, the predicted source-load power data of the corresponding device at ten o'clock can be obtained based on the historical monitoring data at six o'clock, seven o'clock, eight o'clock, and nine o'clock, or the predicted source-load power data of the corresponding device at nine o'clock and ten o'clock can be obtained based on the historical monitoring data at six o'clock, seven o'clock, and eight o'clock. However, the larger the size ratio of the input data set to the predicted source-load power data, the higher the accuracy of the predicted source-load power data.

[0081] Preferably, the steps of inputting the test set into the neural network model to obtain the predicted source-load power data and obtaining the error between the corresponding actual source-load power data to obtain the error probability distribution curve may include:

[0082] S31: Input the test set into the neural network model to obtain the predicted source-load power data of the corresponding device within a set time period;

[0083] S32: Collect the actual source-load power data of the device within the set time period to obtain the prediction error data between it and the predicted source-load power data.

[0084] Specifically, optionally but not limited to the set time period, input the test set in step S1 into the neural network model trained in step S2 to obtain the predicted source-load power data of the corresponding device within the set time period. Then, collect the actual source-load power data of the corresponding device within this set time period to calculate the difference between the actual source-load power data and the predicted source-load power data as the prediction error data.

[0085] Exemplarily, preferably calculate the prediction error data according to Equation 3-1:

[0086]

[0087] Wherein, is the prediction error data, P type,k is the actual source-load power data, is the predicted source-load power data, type represents the energy type, and k is the kth moment in the set time period. The energy type includes any energy type such as wind power and photovoltaic power.

[0088] Preferably, the error between the true data of the source-load power and the predicted data of the source-load power is normalized according to Equation 3-2, making the analysis of the prediction error data more intuitive and convenient:

[0089]

[0090] where ε type,k is the normalized prediction error data, 0 < k ≤ N, N is the number of true data of the source-load power / predicted data of the source-load power, is the maximum value in the prediction error data, and type represents the energy type.

[0091] S33: Resample the prediction error data to obtain a data sample set;

[0092] Specifically, since the Bootstrap method does not require the assumption that the data follows a specific distribution (such as a normal distribution). It estimates the distribution of the statistic by sampling with replacement from the original data and directly using the characteristics of the data itself, and can generate a large number of statistics of resampled samples (such as mean, median, variance, etc.), thereby estimating the distribution of these statistics, and then calculating the confidence interval, standard error, etc. At the same time, the implementation is relatively simple, only requiring sampling with replacement from the original data and calculating the statistic, without complex mathematical derivations. Therefore, it is optional but not limited to using the Bootstrap method to resample the prediction error data to obtain a data sample set.

[0093] Preferably, the steps of resampling the prediction error data to obtain a data sample set may optionally include:

[0094] S331: Resample the prediction error data to obtain a number of data groups;

[0095] Exemplarily, it is optional to sort all the values in the prediction error data in a set order, and then randomly sample with replacement N times to obtain a data group containing N elements as shown in Equation 3-3, and then repeat the above steps M times to obtain M data groups, each data group containing N elements; the set order is arbitrarily set by those skilled in the art.

[0096] Preferably, the data group may optionally be as shown in Equation 3-4:

[0097] {ε type,1 , ε type,2 , ε type,3 ,....... ε type,N} 3-3

[0098]

[0099] where ε type,N is the Nth column element in the data group, is the Nth column element in the ith data group, and type represents the energy type.

[0100] S332: Arrange the elements in each data group in a set order to obtain several updated data groups;

[0101] Exemplarily, optionally arrange the elements in each data group in a set order to obtain several updated data groups as shown in Equation 3-5:

[0102]

[0103] where, is the Nth column element in the ith data group after sorting, 0 < i ≤ M, and M is the number of data groups.

[0104] S333: Obtain the arithmetic mean of the elements in each column of each updated data group to obtain a data sample set.

[0105] Exemplarily, optionally calculate the arithmetic mean of the elements in each column of each updated data group according to Equation 3-6, thereby obtaining a sample data set as shown in Equation 3-7:

[0106]

[0107] where, is the arithmetic mean of the jth column element in the sample data set, 0 < j ≤ N, and N is the number of elements in the data group, is the sample data set.

[0108] S34: Specifically, since the non-parametric test method kernel density estimation can start from the sample data itself and does not require any heuristic knowledge of the data distribution to study the distribution of the sample data. This method has a better fitting effect and can estimate the true distribution of the sample. Therefore, it is preferred to use the kernel density estimation method to obtain the probability distribution of the data set and obtain an error probability distribution curve.

[0109] Preferably, optionally perform kernel density estimation on the sample data set using the kernel density estimation function Equation 3-8 to obtain an error probability distribution curve:

[0110]

[0111] where, f() is the kernel density estimation function, h is the bandwidth, and k() is the selected kernel function, is the arithmetic mean of the jth column element in the sample data set, 0 < j ≤ N, and N is the number of elements in the data group, is the ith element in the sample data set, 0 < i ≤ N, and N is the number of elements in the sample data set. Thus, for each data point, a kernel density function is constructed with this point as the center using the kernel function.

[0112] S4: Randomly sample the error probability distribution curve to obtain prediction error compensation data;

[0113] Specifically, since the Monte Carlo simulation method has high accuracy and a simple calculation mode, it is often used in engineering research and calculations. Preferably, the Monte Carlo simulation method is used to randomly sample the error probability distribution curve to obtain prediction error compensation data, so as to correct the prediction results of the neural network model and make the prediction results closer to the real data.

[0114] Preferably, the steps of randomly sampling the error probability distribution curve to obtain prediction error compensation data may include:

[0115] S41: Randomly sample the error probability distribution curve to obtain a number of error samples;

[0116] S42: Denormalize each error sample to obtain a number of actual error compensation amounts;

[0117] S43: Construct an error compensation data set based on the actual error compensation amounts to obtain prediction error compensation data.

[0118] Specifically, it is optional but not limited to randomly sampling the error probability distribution curve to obtain a number of error samples to simulate the randomness of the prediction error. Since the prediction error is normalized in step S32, each error sample also needs to be denormalized to restore the actual prediction error and obtain a number of actual error compensation amounts. Then, an error compensation data set can be constructed based on the actual error compensation amounts to obtain the prediction error compensation data.

[0119] Exemplarily, it is optional to randomly sample the error probability distribution curve to obtain L error samples {Δε} (l = 1, 2,..., L), and then preferably restore the normalized error sample Δεl to the actual error compensation amount according to Equation 4-1:

[0120]

[0121] where is the l-th error compensation amount, 0 < l ≤ L, and L is the number of error samples.

[0122] Then sum up all the error correction amounts to construct an error compensation data set as the prediction error compensation data for subsequent correction of the prediction data.

[0123] S5: Obtain the final source-load power prediction data based on the prediction set and the prediction error compensation data.

[0124] Specifically, but not limited to, the prediction set constructed based on the device historical monitoring data in step S1 is input into the neural network model trained in step S2 to obtain initial prediction data, and then the initial prediction data is corrected according to the prediction error compensation data obtained in step S4, and the final prediction data of the current device can be obtained.

[0125] Exemplarily, it is optional to obtain the historical power generation data of wind power generation equipment, the historical power generation data of photovoltaic power generation equipment, and the historical load power data. Through steps S1-S5, as well as methods such as the typical day method and the Att-GRU neural network, Figure 4 , Figure 5 and Figure 6 the comparison diagrams of each prediction data as shown can clearly show that the source-load power prediction method based on prediction error correction provided by the present invention can effectively cope with the uncertainty of source-load and improve the prediction accuracy.

[0126] In this embodiment, a source-load power prediction method based on prediction error correction of the present invention is given. By obtaining device historical monitoring data, constructing a training set, a test set, and a prediction set to provide a basis for subsequent steps, and then constructing and training a neural network model with historical monitoring data as input and source-load power prediction data as output according to the training set, so as to obtain the time series characteristics and important information in the prediction set through the neural network model to obtain initial prediction data. Then, according to the test set and the trained neural network model, the error between the prediction data and the real data in the test set is determined to obtain an error probability distribution curve, and then the error probability distribution curve is randomly sampled to obtain prediction error compensation data. By analyzing the prediction error and estimating the probability distribution, data support is provided for subsequent error correction to improve the accuracy of the error distribution. Finally, according to the prediction set and the prediction error compensation data, the final source-load power prediction data is obtained, and the initial prediction data is directly corrected by the prediction error compensation data, so as to effectively cope with the uncertainty of source-load and improve the prediction accuracy. It solves the problems that the prior art cannot effectively cope with the uncertainty of source-load and the prediction accuracy is insufficient.

[0127] On the other hand, the present invention also provides a computer storage medium storing executable program code; the executable program code is used to execute any of the above-mentioned source-load power prediction methods based on prediction error correction.

[0128] On the other hand, the present invention also provides a terminal device including a memory and a processor; the memory stores program code executable by the processor; the program code is used to execute any of the above-mentioned source-load power prediction methods based on prediction error correction.

[0129] Exemplarily, the program code can be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the program code in the terminal device.

[0130] The terminal device can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The terminal device may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the terminal device may further include input / output devices, network access devices, a bus, etc.

[0131] The processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0132] The memory may be an internal storage unit of the terminal device, such as a hard disk or a memory. The memory may also be an external storage device of the terminal device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal device. Further, the memory may also include both the internal storage unit and the external storage device of the terminal device. The memory is used to store the program code and other programs and data required by the terminal device. The memory may also be used to temporarily store the data that has been output or is to be output.

[0133] The above computer storage medium and terminal device are created based on the above source-load power prediction method based on prediction error correction. Their technical effects and beneficial effects are not elaborated herein. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope described in this specification.

[0134] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.

Claims

1. A source-load power prediction method based on prediction error correction, characterized in that Including: Obtain the historical monitoring data of the device, and construct a training set, a test set, and a prediction set; Construct and train a neural network model with the historical monitoring data as the input and the source-load power prediction data as the output according to the training set; According to the test set and the trained neural network model, determine the error between the predicted data and the real data in the test set, and obtain the error probability distribution curve; Perform random sampling on the error probability distribution curve to obtain the prediction error compensation data; According to the prediction set and the prediction error compensation data, obtain the final source-load power prediction data.

2. The method according to claim 1, wherein The neural network model includes: An input layer for receiving the historical monitoring data of the device; A feature extraction layer connected to the input layer, which is used to generate n input vector sequences according to the historical monitoring data and extract the temporal features of each input vector sequence; n is the number of time nodes of the historical monitoring data; A feature fusion layer connected to the feature extraction layer, which is used to configure weights for each temporal feature to generate weighted vectors; An output layer connected to the feature fusion layer, which is used to output the source-load power prediction data according to the weighted vectors.

3. The method according to claim 2, wherein The feature extraction layer includes: A generation unit for generating n input vector sequences according to the historical monitoring data; And n GRU units connected in chronological order; the GRU unit at time t has two input vectors: the input vector sequence at the current moment and the output of the previous GRU unit; it is used to generate a hidden state sequence as the temporal feature of the current input vector according to the input vector sequence.

4. The method according to claim 3, characterized in that, The generation unit includes: A division sub-unit connected to the input layer, which divides the historical monitoring data into input vector sequences at m moments according to the time node interval; m = n / 2; An analysis sub-unit connected to the output of the division sub-unit, which is used to analyze the day-night data of each input vector sequence, divide each input vector sequence into a day input vector sequence and a night input vector sequence, obtain n input vector sequences, and input them into n GRU layers.

5. The method according to claim 3, characterized in that, The feature fusion layer is an attention layer connected to the GRU layer, which is used to perform weighted processing on the hidden state sequence output by the GRU layer to obtain attention weights, so as to generate weighted vectors according to the hidden state sequence and the attention weights.

6. The method according to claim 1, wherein The steps to obtain the error probability distribution curve include: Input the test set into the neural network model to obtain the source-load power prediction data of the corresponding device within a set time period; Collect the real source-load power data of the device within the set time period to obtain the prediction error data between it and the source-load power prediction data; Resample the prediction error data to obtain a data sample set; Perform kernel density estimation on the data sample set to obtain the error probability distribution curve.

7. The method according to claim 6, characterized in that, The steps to resample the prediction error data to obtain a data sample set include: Resample the prediction error data to obtain several data groups; Arrange the elements in each data group in a set order to obtain several updated data groups; Obtain the arithmetic mean of the elements in each column of the updated data groups to obtain the data sample set.

8. The method according to claim 1, characterized in that, The steps to perform random sampling on the error probability distribution curve to obtain the prediction error compensation data include: Perform random sampling on the error probability distribution curve to obtain several error samples; Denormalize each error sample to obtain a number of actual error compensation amounts; Construct an error compensation data set based on the actual error compensation amounts to obtain predicted error compensation data.

9. A computer storage medium, characterized in that, Store executable program code; the executable program code is used to execute the source-load power prediction method based on predicted error correction according to any one of claims 1-8.

10. A terminal device, characterized in that, Comprise a memory and a processor; the memory stores program code executable by the processor; the program code is used to execute the source-load power prediction method based on predicted error correction according to any one of claims 1-8.

Images