A power prediction method, system and storage medium for power supply type virtual power plant

Abstract

The present invention discloses a power prediction method, system and storage medium for a power type virtual power plant in the technical field of power rate prediction of a power type virtual power plant. The method comprises preprocessing the original power sequence of the power type virtual power plant; decomposing the preprocessed original power sequence to obtain the principal component sequence under long and short time scales; using a wide network to obtain the mapping relationship between the power data of the power type virtual power plant to be predicted under a long time scale and the input principal component sequence, and outputting the prediction result; using a deep network to obtain the mapping relationship between the power data of the power type virtual power plant to be predicted under a short time scale and the input principal component sequence, and outputting the prediction result; using a fully connected layer to fuse the prediction result of the wide network with the prediction result of the deep network to obtain the final power prediction result of the power type virtual power plant. The present invention can reduce the power prediction error of the power type virtual power plant and improve the accuracy and stability of the prediction result.

Description

Technical Field

[0001] The present invention belongs to the technical field of power rate prediction of power supply type virtual power plants, and in particular relates to a power prediction method, system and storage medium for power supply type virtual power plants. Background Art

[0002] Power-based virtual power plants organically combine distributed power sources, controllable loads, energy storage systems, electric vehicles, etc., and participate in the power market and power grid operation as a special power plant, which can safely and efficiently utilize large-scale new energy power. However, due to the wide variety of energy types aggregated by power-based virtual power plants, new energy sources such as wind power and photovoltaic power are intermittent and volatile, which poses challenges to the grid-connected operation of power-based virtual power plants. By predicting the power of power-based virtual power plants, the balance of distributed energy supply and demand can be achieved, providing important guarantees for the grid-connected consumption of virtual power plants.

[0003] The widely used power prediction methods for power-based virtual power plants are mainly divided into statistical methods and physical methods, and also include combined prediction methods that combine multiple methods. Among them, the physical method is a prediction method based on virtual power plant power historical data and multi-factor virtual power plant physical data such as numerical weather forecasts. The advantages of virtual power plant power prediction based on physical methods are that the amount of data relied on is small, the model is simple, and the prediction is convenient and fast; but there is a large error between the predicted physical data and the real data, which makes the physical method generally low in wind power prediction accuracy. The statistical method is a virtual power plant power prediction method that updates and adjusts the model weights and parameters based on the error between the predicted value and the real value. This method is usually divided into probabilistic statistical models, machine learning models, and deep learning models. Compared with the physical method, the probabilistic statistical model can better follow the changing trend of virtual power plant power; the prediction accuracy of the machine learning model is better than that of the traditional probabilistic model; the deep learning model can fully explore the time and space characteristics of the input sequence, and further improve the accuracy of virtual power plant power prediction. The statistical method can well extract the linear and nonlinear relationship of historical data and improve the prediction accuracy of virtual power plant power, but it has high requirements for the consistency of the change law of historical data. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, the present invention provides a power prediction method, system and storage medium for a power supply type virtual power plant, which can reduce the power prediction error of the power supply type virtual power plant and improve the accuracy and stability of the prediction results.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] In the first aspect, a power prediction method for a power supply type virtual power plant based on a wide and deep neural network is provided, comprising: preprocessing the original power sequence of the power supply type virtual power plant; decomposing the preprocessed original power sequence to obtain the principal component sequences at long and short time scales respectively; using a wide network TCN to obtain a mapping relationship between the power data of the power supply type virtual power plant to be predicted at a long time scale and the input principal component sequence, and outputting the prediction result of the wide network TCN; using a deep network LSTM to obtain a mapping relationship between the power data of the power supply type virtual power plant to be predicted at a short time scale and the input principal component sequence, and outputting the prediction result of the deep network LSTM; using a fully connected layer to fuse the prediction result of the wide network TCN with the prediction result of the deep network LSTM to obtain the final power prediction result of the power supply type virtual power plant.

[0007] Furthermore, the preprocessing of the original power sequence of the power supply type virtual power plant includes: processing negative values ​​by directly setting to zero, correcting overfull values ​​and garbled codes by using the power values ​​of the power supply type virtual power plant at two moments before and after, filling in the missing values, and normalizing the original power sequence. The normalization expression is:

[0008]

[0009] Among them, P t ” is the power of the power type virtual power plant at time t after normalization, P t ' is the power of the power type virtual power plant at time t before normalization, max(P') is the maximum value in the power type virtual power plant power sequence after preprocessing, and min(P') is the minimum value in the power type virtual power plant power sequence after preprocessing.

[0010] Furthermore, the decomposition of the preprocessed original power sequence to obtain the principal component sequences at long and short time scales respectively includes: constraining the variation of the preprocessed original power sequence; converting the constrained variation result into an unconstrained variation problem by introducing Lagrange multipliers and quadratic penalty functions; optimizing the unconstrained variation problem by using the alternating direction multiplier method iterative algorithm combined with Parseval, Plancher theorems and Fourier isometries, and obtaining the modal components and center frequencies of the principal component sequences at long and short time scales respectively.

[0011] Furthermore, the constraint condition of the constrained variation is that the sum of all modes is equal to the original signal, the decomposition sequence is a modal component with a finite bandwidth and a center frequency, and the sum of the estimated bandwidths of each mode is minimal. The constrained variation expression is:

[0012]

[0013] Among them, u m={u1,u2,...,u m} is the modal components of the power sequence of the power supply type virtual power plant, is the square expression of the two-norm, δ(t) is the impulse function, θ t is the first-order partial derivative with respect to time t, f(t) is the original power sequence of the power supply type virtual power plant, m is the number of components at different time scales, m is 200 dimensions at a long time scale, and m is 96 dimensions at a short time scale. is the Fourier transform basis.

[0014] Furthermore, the expression of the unconstrained variational problem is:

[0015]

[0016] Among them, u m (t)={u1(t),u2(t),...,u m (t)} is the modal components of the power sequence of the power supply type virtual power plant decomposed at time t, ω m ={ω1,ω2,...,ω m} is the center frequency of each modal component, is the square expression of the two-norm, δ(t) is the impulse function, θ t is the first-order partial derivative with respect to time t, f(t) is the original power sequence of the power supply type virtual power plant, m is the number of components at different time scales, m is 200 dimensions at long time scales, and m is 96 dimensions at short time scales, α is the introduced quadratic penalty factor parameter, and λ(t) is the Lagrange multiplier at time t.

[0017] Furthermore, the expressions of the modal components and center frequencies of the principal component sequences under the long and short time scales are as follows:

[0018]

[0019]

[0020]

[0021] Among them, τ is the noise tolerance, which satisfies the fidelity of power sequence decomposition of power supply type virtual power plant, and u m s+1 (t),u i The Fourier transform of (t), f(t) and λ(t), is the L2 distance between the current frequency and the center frequency obtained in the sth iteration, α is the introduced quadratic penalty factor parameter, is the center frequency of each modal component after the s+1th iteration, is the Lagrange multiplier after the s+1th iteration, f(t) is the original power sequence of the power supply type virtual power plant, u i (t) is the i-th modal component at time t, are the modal components after the s+1th iteration, and λ(t) is the Lagrange multiplier at time t.

[0022] Furthermore, the use of the wide network TCN to obtain the mapping relationship between the power data of the power type virtual power plant to be predicted and the input principal component sequence at a long time scale includes: the output sequence Ps obtained after constrained variation of the input principal component sequence, mining the mapping relationship between the virtual power plant power and the principal component sequence at a long time scale by stacking hole convolution, and outputting the power value P of the power type virtual power plant at the next moment t , where the power value of the power supply type virtual power plant at time t is P t The dilated convolution is:

[0023]

[0024] Where S = {s1,s2,...,s T} is the input sequence, s i , i∈[1,T] is a column vector, d represents the dilation factor, K represents the filter size, and the filter is F={f1,f2,…,f K}, * is the convolution symbol, s t-d(K-k) is the td(Kk)th input sequence vector s i , the value range of i is [tK,t].

[0025] Furthermore, the use of the deep network LSTM to obtain the mapping relationship between the power data of the power type virtual power plant to be predicted under a short time scale and the input principal component sequence includes: mining the mapping relationship between the power data of the virtual power plant to be predicted under a short time scale and the input principal component sequence through memory state update, and outputting the power value P of the power type virtual power plant at the next moment. t , where the power value of the power supply type virtual power plant at time t is P t The status is updated to:

[0026]

[0027] Among them, θ is the sigmoid activation function, φ is the tanh activation function; W fp is the forget gate and the input power value P of the power type virtual power plant at the next moment t The multiplied weight matrix, W fh is the forget gate and the previous hidden state Ht-1 The multiplied weight matrix, W ip The power value P of the power supply type virtual power plant at the next moment is input to the gate t The multiplied weight matrix, W ih is the input gate and the previous hidden state H t-1 The multiplied weight matrix, W gp The power value P of the power supply type virtual power plant at the next moment is the input node t The multiplied weight matrix, W gh is the input node and the previous hidden state H t-1 The multiplied weight matrix, W op The power value P of the power supply type virtual power plant at the next moment between the output gate and the input t The multiplied weight matrix, W oh is the output gate and the previous hidden state H t-1 The multiplied weight matrix; e f is the bias term corresponding to the forget gate, e i is the bias term corresponding to the input gate, e g is the bias term corresponding to the input node, e0 is the bias term corresponding to the output gate; f t is the output result of the forget gate, i t is the output result of the input gate, g t is the output result of the input node, o t is the output result of the output gate, C t is the output result of the memory unit state, H t is the output result of the hidden state.

[0028] Furthermore, the prediction result of the wide network TCN is integrated with the prediction result of the deep network LSTM by using the fully connected layer, including: the prediction result of the wide network TCN is integrated with the prediction result of the deep network LSTM and then linearized and outputted to obtain the predicted power of the power supply type virtual power plant:

[0029] α t =W[O TCN ,O LSTM ]+b (9)

[0030] Among them, α t is the linearized output after fusion of the fully connected layer, W is the weight matrix of the fully connected layer, O TCN is the prediction result of wide network TCN, O LSTM is the prediction result of the deep network LSTM, and b is the bias matrix of the fully connected layer.

[0031] In the second aspect, a power prediction system for a power supply type virtual power plant based on a wide and deep neural network is provided, including: a preprocessing module, used to preprocess the original power sequence of the power supply type virtual power plant; a decomposition module, used to decompose the preprocessed original power sequence to obtain the principal component sequences under long and short time scales respectively; a first data processing module, used to use the wide network TCN to obtain the mapping relationship between the power data of the power supply type virtual power plant to be predicted and the input principal component sequence under long time scale, and output the prediction result of the wide network TCN; a second data processing module, used to use the deep network LSTM to obtain the mapping relationship between the power data of the power supply type virtual power plant to be predicted and the input principal component sequence under short time scale, and output the prediction result of the deep network LSTM; a prediction module, used to use the fully connected layer to fuse the prediction result of the wide network TCN with the prediction result of the deep network LSTM to obtain the final power prediction result of the power supply type virtual power plant.

[0032] According to a third aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the method described in the first aspect.

[0033] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: the present invention obtains the mapping relationship between the power data of the power supply type virtual power plant to be predicted and the input principal component sequence under a long time scale by using the wide network TCN; obtains the mapping relationship between the power data of the power supply type virtual power plant to be predicted and the input principal component sequence under a short time scale by using the deep network LSTM; uses a fully connected layer to fuse the prediction result of the wide network TCN with the prediction result of the deep network LSTM to obtain the final power prediction result of the power supply type virtual power plant; can effectively predict the total power of distributed energy of the power supply type virtual power plant, can reduce the power prediction error of the power supply type virtual power plant, and improve the accuracy and stability of the prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of a power prediction method for a power supply type virtual power plant based on a wide and deep neural network provided by an embodiment of the present invention;

[0035] Figure 2 is a dilated convolutional graph of a wide network TCN model in an embodiment of the present invention;

[0036] Figure 3 is a memory neural graph of a deep network LSTM module in an embodiment of the present invention;

[0037] Figure 4It is a power prediction curve of a power supply type virtual power plant based on a wide and deep neural network obtained in an embodiment of the present invention, wherein (a) is a virtual power plant power prediction value obtained by a TCN network LSTM network and a wide and deep TCN-LSTM network at a time scale of 15 minutes, (b) is a virtual power plant power prediction value obtained by a TCN network LSTM network and a wide and deep TCN-LSTM network at a time scale of 30 minutes, (c) is a virtual power plant power prediction value obtained by a TCN network LSTM network and a wide and deep TCN-LSTM network at a time scale of 45 minutes, and (d) is a virtual power plant power prediction value obtained by a TCN network LSTM network and a wide and deep TCN-LSTM network at a time scale of 60 minutes. DETAILED DESCRIPTION

[0038] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0039] Embodiment 1:

[0040] like Figure 1 As shown, a power prediction method for a power supply type virtual power plant based on a wide and deep neural network includes: preprocessing the original power sequence of the power supply type virtual power plant; decomposing the preprocessed original power sequence to obtain the principal component sequences under long and short time scales respectively; using a wide network TCN to obtain a mapping relationship between the power data of the power supply type virtual power plant to be predicted and the input principal component sequence under a long time scale, and outputting the prediction result of the wide network TCN; using a deep network LSTM to obtain a mapping relationship between the power data of the power supply type virtual power plant to be predicted and the input principal component sequence under a short time scale, and outputting the prediction result of the deep network LSTM; using a fully connected layer to fuse the prediction result of the wide network TCN with the prediction result of the deep network LSTM to obtain the final power prediction result of the power supply type virtual power plant.

[0041] The specific implementation process of this embodiment is as follows.

[0042] Step 1: Preprocess the original power sequence of the power supply type virtual power plant; the preprocessing includes abnormal value processing (negative value, power value exceeding full load and garbled code) and filling of vacant values. The negative value is processed by directly setting it to zero, and the power value of the power supply type virtual power plant at the two moments before and after is used to correct or fill the over-full power value, garbled code and vacant value. Considering that the installed capacity of various types of generators in the virtual power plant is different, resulting in differences in their power generation, the experimental data is normalized. The specific formula is:

[0043]

[0044] Among them, P t” is the power of the power type virtual power plant at time t after normalization, P t ' is the power of the power type virtual power plant at time t before normalization, max(P') is the maximum value in the power type virtual power plant power sequence after preprocessing, and min(P') is the minimum value in the power type virtual power plant power sequence after preprocessing.

[0045] Step 2: Use the VMD (Variational Mode Decomposition) module to decompose the power sequence of the power supply type virtual power plant and obtain the principal component sequences at long and short time scales, including:

[0046] The constrained variation is performed on the preprocessed original power sequence; the constraint condition of the constrained variation is that the sum of all modes is equal to the original signal, the decomposed sequence is a modal component with a finite bandwidth and a center frequency, and the sum of the estimated bandwidths of each mode is minimized. The constrained variation expression is:

[0047]

[0048] Among them, u m ={u1,u2,...,u m} is the modal components of the power sequence of the power supply type virtual power plant, is the square expression of the two-norm, δ(t) is the impulse function, θ t is the first-order partial derivative with respect to time t, f(t) is the original power sequence of the power supply type virtual power plant, m is the number of components at different time scales, m is 200 dimensions at a long time scale, and m is 96 dimensions at a short time scale. is the Fourier transform basis.

[0049] By introducing Lagrange multipliers and quadratic penalty functions, the constrained variational results are transformed into unconstrained variational problems; the expression of the unconstrained variational problem is:

[0050]

[0051] Among them, u m (t)={u1(t),u2(t),...,u m (t)} is the modal components of the power sequence of the power supply type virtual power plant decomposed at time t, ω m ={ω1,ω2,...,ω m} is the center frequency of each modal component, is the square expression of the two-norm, δ(t) is the impulse function, θ t is the first-order partial derivative with respect to time t, f(t) is the original power sequence of the power supply type virtual power plant, m is the number of components at different time scales, m is 200 dimensions at long time scales, and m is 96 dimensions at short time scales, α is the introduced quadratic penalty factor parameter, and λ(t) is the Lagrange multiplier at time t;

[0052] The alternating direction multiplier iterative algorithm is combined with Parseval, Plancherel theorem and Fourier isometric transform to optimize the unconstrained variational problem, and the modal components and center frequencies of the principal component sequence under long and short time scales are obtained respectively; the expressions of the modal components and center frequencies of the principal component sequence under long and short time scales are:

[0053]

[0054]

[0055]

[0056] Among them, τ is the noise tolerance, which satisfies the fidelity of power sequence decomposition of power supply type virtual power plant, and u m s+1 (t),u i The Fourier transform of (t), f(t) and λ(t), is the L2 distance between the current frequency and the center frequency obtained in the sth iteration, α is the introduced quadratic penalty factor parameter, is the center frequency of each modal component after the s+1th iteration, is the Lagrange multiplier after the s+1th iteration, f(t) is the original power sequence of the power supply type virtual power plant, u i (t) is the i-th modal component at time t, u m s+1 (t) is the modal component after the s+1th iteration, and λ(t) is the Lagrange multiplier at time t.

[0057] Step 3: Use the wide network TCN to obtain the mapping relationship between the power data of the power type virtual power plant to be predicted and the input principal component sequence on a long time scale, and output the prediction result of the wide network TCN;

[0058] The output sequence Ps obtained after constrained variation of the input principal component sequence is used to mine the mapping relationship between the virtual power plant power and the principal component sequence at a long time scale through stacked dilated convolution, and the power value of the power supply type virtual power plant at the next moment is output;

[0059] Let filter F = (f1, f2, ..., f K ), the output sequence information is Y = (y1, y2, ..., y T ), where the input is a stationary sequence with symmetric center frequency obtained after VMD decomposition: Ps = {P1, P2, ..., P T}, where Pi , i∈[1,T] is a column vector; the following formula gives the power P of the power type virtual power plant at time t: t Atrous convolution:

[0060] Among them, the power of the power type virtual power plant at time t is P t The dilated convolution is:

[0061]

[0062] Where S = {s1,s2,...,s T} is the input sequence, s i , i∈[1,T] is a column vector, d represents the dilation factor, K represents the filter size, and the filter is F={f1,f2,…,f K}, * is the convolution symbol, s t-d(K -k) is the td(Kk)th input sequence vector s i , the value range of i is [tK,t].

[0063] like Figure 2 As shown in the figure, the wide network TCN model horizontally expands the original neural network by increasing the number of neural nodes. The network replaces the direct connection between input and output with the mapping relationship between the virtual power plant power data to be predicted and the input principal component sequence over a long time scale, thereby improving the accuracy of the model.

[0064] Step 4: Use the deep network LSTM to obtain the mapping relationship between the power data of the power type virtual power plant to be predicted in a short time scale and the input principal component sequence, and output the prediction result of the deep network LSTM;

[0065] The output sequence Ps obtained after constrained variation of the input principal component sequence is used to mine the mapping relationship between the power data of the virtual power plant to be predicted under short time scale and the input principal component sequence through memory state update, and the power value of the power supply type virtual power plant at the next moment is output;

[0066] LSTM updates and regulates the signal state information through the forget gate, input layer gate and output layer gate, discards worthless information, and retains and fills in useful information. t The status is updated to:

[0067]

[0068] Among them, θ is the sigmoid activation function, φ is the tanh activation function; W fp is the forget gate and the input power value P of the power type virtual power plant at the next moment t The multiplied weight matrix, W fh is the forget gate and the previous hidden state Ht-1 The multiplied weight matrix, W ip The power value P of the power supply type virtual power plant at the next moment is input to the gate t The multiplied weight matrix, W ih is the input gate and the previous hidden state H t-1 The multiplied weight matrix, W gp The power value P of the power supply type virtual power plant at the next moment is the input node t The multiplied weight matrix, W gh is the input node and the previous hidden state H t-1 The multiplied weight matrix, W op The power value P of the power supply type virtual power plant at the next moment between the output gate and the input t The multiplied weight matrix, W oh is the output gate and the previous hidden state H t-1 The multiplied weight matrix; e f is the bias term corresponding to the forget gate, e i is the bias term corresponding to the input gate, e g is the bias term corresponding to the input node, e0 is the bias term corresponding to the output gate; f t is the output result of the forget gate, i t is the output result of the input gate, g t is the output result of the input node, o t is the output result of the output gate, C t is the output result of the memory unit state, H t is the output result of the hidden state.

[0069] like Figure 3 As shown in the figure, the deep network LSTM increases the number of layers of the network structure based on the original neural network, deeply mines the mapping relationship between the virtual power plant power data to be predicted and the input principal component sequence under short time scales, and can calculate more complex input features to improve the accuracy of the model.

[0070] Step 5: Use the fully connected layer to fuse the prediction results of the wide network TCN with the prediction results of the deep network LSTM to obtain the final power prediction results of the power supply type virtual power plant; fuse the prediction results of the wide network TCN with the prediction results of the deep network LSTM and then linearize the output to obtain the predicted power of the power supply type virtual power plant:

[0071] α t =W[O TCN ,O LSTM ]+b (9)

[0072] Among them, α t is the linearized output after fusion of the fully connected layer, W is the weight matrix of the fully connected layer, O TCNis the prediction result of wide network TCN, O LSTM is the prediction result of the deep network LSTM, and b is the bias matrix of the fully connected layer.

[0073] The prediction results of different prediction models are evaluated for performance. In this embodiment, multiple performance indicators are established to evaluate the prediction results, including mean absolute error (MAE), root mean square error (RMSE), MAE improvement of model prediction results (P MAE ) and RMSE improvement (P RMSE ), the specific formula is as follows:

[0074]

[0075]

[0076]

[0077]

[0078] Among them, y i and are the true value and predicted value at time i respectively; Model MAE and Model RMSE They are the MAE and RMSE of the models, where Model1 is the combined model in this embodiment and Model2 is the comparison model.

[0079] Using the method of this embodiment, the field measured data of a virtual power plant in a city in Jiangsu Province from December 2020 to May 2021 are used, with a sampling interval of 15 minutes. There are more than 15,000 data in total, and 70% of them are used as training data sets after shuffling, and 30% are used as test data sets. Figure 4 The power prediction curve of power supply type virtual power plant based on wide and deep learning network was obtained.

[0080] In order to verify the superiority of the proposed wide and deep TCN-LSTM model in power prediction of power-based virtual power plants, the power prediction performance of TCN, LSTM and wide and deep TCN-LSTM models at three different time scales (15min, 30min, 1h) is analyzed and compared. The prediction results are shown in Table 1:

[0081] Table 1

[0082]

[0083]

[0084] As shown in Table 1, it can be seen that the wide and deep TCN-LSTM model performs better than the other two models in terms of MAE and MRE indicators. The wide network TCN model horizontally expands the original neural network by increasing the number of neural nodes. The network replaces the direct connection between input and output with the mapping relationship between the virtual power plant power data to be predicted and the input principal component sequence on a long time scale; the deep network LSTM increases the number of layers of the network structure on the basis of the original neural network, deeply mines the mapping relationship between the virtual power plant power data to be predicted and the input principal component sequence on a short time scale, and can calculate more complex input features. By combining the wide network TCN and the deep network LSTM, it can have a more efficient learning performance, which is more conducive to improving the prediction accuracy of the power of the power supply type virtual power plant.

[0085] This embodiment uses variational mode decomposition (VMD) and wide network TCN to extract the mapping relationship between the power data of the power supply type virtual power plant to be predicted and the input principal component sequence under long time scales, and outputs the prediction results of the wide network; then uses variational mode decomposition (VMD) and deep network LSTM to deeply mine the mapping relationship between the power data of the virtual power plant to be predicted and the input principal component sequence under short time scales, and outputs the prediction results of the deep network. Finally, the prediction results of the wide and deep learning network are fused using a fully connected layer to obtain the final power prediction results of the power supply type virtual power plant. The comparison of different prediction models verifies that the proposed model effectively improves the accuracy and stability of the prediction.

[0086] The power prediction method for a power supply type virtual power plant based on a wide and deep learning network provided in this embodiment, the VMD can decompose to obtain a stable signal component with a symmetrical center frequency that is more conducive to prediction, can effectively reduce the impact of the power noise of the original power supply type virtual power plant on the prediction, and improve the prediction accuracy; the combined prediction model of the wide network TCN and the deep network LSTM can effectively improve the accuracy and stability of the power prediction of the power supply type virtual power plant.

[0087] Embodiment 2:

[0088] Based on the power prediction method of a power supply type virtual power plant based on a wide and deep neural network described in the first embodiment, this embodiment provides a power prediction system of a power supply type virtual power plant based on a wide and deep neural network, including:

[0089] A preprocessing module, used for preprocessing the original power sequence of the power supply type virtual power plant;

[0090] The decomposition module is used to decompose the original power sequence after preprocessing to obtain the principal component sequences at long and short time scales respectively;

[0091] The first data processing module is used to use the wide network TCN to obtain the mapping relationship between the power data of the power type virtual power plant to be predicted and the input principal component sequence at a long time scale, and output the prediction result of the wide network TCN;

[0092] The second data processing module is used to use the deep network LSTM to obtain the mapping relationship between the power data of the power type virtual power plant to be predicted and the input principal component sequence in a short time scale, and output the prediction result of the deep network LSTM;

[0093] The prediction module is used to use the fully connected layer to fuse the prediction results of the wide network TCN with the prediction results of the deep network LSTM to obtain the final power prediction results of the power supply type virtual power plant.

[0094] Embodiment three:

[0095] Based on the power prediction method of a power supply type virtual power plant based on a wide and deep neural network described in Example 1, and the power prediction system of a power supply type virtual power plant based on a wide and deep neural network described in Example 2, this embodiment provides a computer-readable storage medium, and the computer-readable storage medium includes a stored computer program, wherein when the computer program is executed by the processor, the device where the storage medium is located is controlled to execute the method described in Example 1.

[0096] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0097] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0098] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0099] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

Claims

1. A power prediction method for a power supply type virtual power plant based on a wide and deep neural network, characterized in that: include: Preprocessing the original power sequence of the power supply type virtual power plant; Decompose the original power sequence after preprocessing to obtain the principal component sequences at long and short time scales respectively; Using the wide network TCN, the mapping relationship between the power data of the power type virtual power plant to be predicted and the input principal component sequence in a long time scale is obtained, and the prediction result of the wide network TCN is output; Using the deep network LSTM, the mapping relationship between the power data of the power type virtual power plant to be predicted in a short time scale and the input principal component sequence is obtained, and the prediction result of the deep network LSTM is output; The prediction results of the wide network TCN and the deep network LSTM are combined using the fully connected layer to obtain the final power prediction results of the power supply type virtual power plant. The preprocessed original power sequence is decomposed to obtain the main component sequences at long and short time scales, respectively, including: Perform constrained variation on the preprocessed original power series; By introducing Lagrange multipliers and quadratic penalty functions, the constrained variational results are transformed into unconstrained variational problems. The alternating direction multiplier iterative algorithm is combined with Parseval's theorem, Plancher's theorem and Fourier isometric transform to optimize the unconstrained variational problem, and the modal components and central frequencies of the principal component sequence at long and short time scales are obtained respectively. The constraint conditions of the constrained variation are that the sum of all modes is equal to the original signal, the decomposition sequence is a modal component with a finite bandwidth and a center frequency, and the sum of the estimated bandwidths of each mode is minimal. The constrained variation expression is: （2） in, To decompose the modal components of the power sequence of the power supply type virtual power plant, is the square expression of the two-norm, is the impulse function, For time The first-order partial derivative of is the original power sequence of the power supply type virtual power plant, is the number of components at different time scales, and 200 dimensions, short time scale is 96 dimensions, is the Fourier transform basis; The expression of the unconstrained variational problem is: （3） in, for Decompose the modal components of the power sequence of the power supply type virtual power plant at all times, is the center frequency of each modal component, is the square expression of the two-norm, is the impulse function, For time The first-order partial derivative of is the original power sequence of the power supply type virtual power plant, is the number of components at different time scales, and 200 dimensions, short time scale is 96 dimensions, is the introduced quadratic penalty factor parameter, for Lagrange multiplier at time; The expressions of the modal components and center frequencies of the principal component sequences under the long and short time scales are: （4） （5） （6） in, is the noise tolerance and satisfies the fidelity of power sequence decomposition of power supply type virtual power plant. , , and They are , , and The Fourier transform of The current frequency and The center frequency obtained by the iteration distance, is the introduced quadratic penalty factor parameter, For the The center frequency of each modal component after iterations is: For the After iterations, the Lagrange multiplier is: is the original power sequence of the power supply type virtual power plant, for Moment modal components, For the After iterations, the modal components are: for Lagrange multiplier at time; The method uses a wide network TCN to obtain a mapping relationship between the power data of the power supply type virtual power plant to be predicted and the input principal component sequence at a long time scale, including: constraining the input principal component sequence to obtain an output sequence By stacking dilated convolutions, the mapping relationship between the virtual power plant power and the principal component sequence at a long time scale is mined, and the power value of the power supply type virtual power plant at the next moment is output. ,in, Power value of virtual power plant with power supply at all times The dilated convolution is: （7） in, is the input sequence, is a column vector, represents the expansion factor, Represents the filter size, the filter is , is the convolution symbol, For the Input sequence vector , The value range is ; The method uses the deep network LSTM to obtain the mapping relationship between the power data of the power supply type virtual power plant to be predicted under a short time scale and the input principal component sequence, including: the output sequence obtained by constraining the input principal component sequence By updating the memory state, the mapping relationship between the power data of the virtual power plant to be predicted and the input principal component sequence in a short time scale is mined, and the power value of the power type virtual power plant at the next moment is output. ,in, Power value of virtual power plant with power supply at all times The status is updated to: （8） in, is the sigmoid activation function, is the tanh activation function; The forget gate and the power value of the power supply type virtual power plant at the next moment are input The multiplied weight matrix, is the forget gate and the previous hidden state The multiplied weight matrix, The power value of the power supply type virtual power plant at the next moment is the input gate and the input The multiplied weight matrix, is the input gate and the previous hidden state The multiplied weight matrix, The power value of the power supply type virtual power plant at the next moment is the input node The multiplied weight matrix, is the input node and the previous hidden state The multiplied weight matrix, The power value of the power supply type virtual power plant at the next moment between the output gate and the input The multiplied weight matrix, is the output gate and the previous hidden state The multiplied weight matrix; is the bias term corresponding to the forget gate, is the bias term corresponding to the input gate, is the bias term corresponding to the input node, is the bias term corresponding to the output gate; is the output of the forget gate, is the output result of the input gate, is the output result of the input node, is the output result of the output gate, is the output result of the memory cell state, is the output result of the hidden state; The method of using a fully connected layer to fuse the prediction result of the wide network TCN with the prediction result of the deep network LSTM includes: fusing the prediction result of the wide network TCN with the prediction result of the deep network LSTM and then linearizing and outputting them to obtain the predicted power of the power supply type virtual power plant: （9） in, is the linearized output after fusion of the fully connected layer, is the weight matrix of the fully connected layer, is the prediction result of wide network TCN, is the prediction result of the deep network LSTM, is the bias matrix of the fully connected layer.

2. The power prediction method of power supply type virtual power plant based on wide and deep neural network according to claim 1 is characterized in that: The preprocessing of the original power sequence of the power supply type virtual power plant includes: processing negative values ​​by directly setting to zero, correcting the over-full value and garbled code by using the power values ​​of the power supply type virtual power plant at two moments before and after, filling the vacant values, and normalizing the original power sequence. The normalization expression is: （1） in, After normalization The power of the power supply type virtual power plant at the moment, Before normalization The power of the power supply type virtual power plant at the moment, The maximum value in the power sequence of the preprocessed power supply type virtual power plant, The minimum value in the power sequence of the preprocessed power type virtual power plant.

3. A power prediction system for a power supply type virtual power plant based on a wide and deep neural network, characterized in that: A method for predicting power of a power supply type virtual power plant based on a wide and deep neural network according to any one of claims 1 to 2; the system comprises: A preprocessing module, used for preprocessing the original power sequence of the power supply type virtual power plant; The decomposition module is used to decompose the original power sequence after preprocessing to obtain the principal component sequences at long and short time scales respectively; The first data processing module is used to use the wide network TCN to obtain the mapping relationship between the power data of the power type virtual power plant to be predicted and the input principal component sequence at a long time scale, and output the prediction result of the wide network TCN; The second data processing module is used to use the deep network LSTM to obtain the mapping relationship between the power data of the power type virtual power plant to be predicted and the input principal component sequence in a short time scale, and output the prediction result of the deep network LSTM; The prediction module is used to use the fully connected layer to fuse the prediction results of the wide network TCN with the prediction results of the deep network LSTM to obtain the final power prediction results of the power supply type virtual power plant.

4. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the method according to any one of claims 1 to 2.

Images