Bidirectional prediction based dispatch optimization method and system for distributed energy resources
By performing orthogonal wavelet transformation and feature extraction on historical data, combining bidirectional prediction methods to integrate environmental and load information, and constructing a distributed energy optimization scheduling model, it solves the problem that traditional scheduling methods are difficult to adapt to the characteristics of distributed energy, and improves prediction accuracy and scheduling efficiency.
Patent Information
- Application Number
- PCT/CN2023/101611
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-16
- Filing Date
- 2023-06-21
- Publication Date
- 2024-11-21
AI Technical Summary
The traditional distribution network scheduling method is difficult to adapt to the randomness, volatility and intermittentness of distributed energy sources such as wind and photovoltaics, which leads to the inability to effectively implement the scheduling strategy, and the existing prediction methods are slow or have poor results when predicting long-term series.
By obtaining historical data, orthogonal wavelet transformation and reconstruction, combining prediction models of global features and mutation characteristics for feature extraction and weighting fusion, two-way prediction methods are used to fuse environment, load and scheduling information, and build a distributed energy optimization scheduling model to improve Prediction accuracy.
The accuracy of distributed energy output prediction and the optimization of scheduling scheme are improved, making the overall scheduling of massive distributed energy more reasonable and efficient.
Smart Images

Figure CN2023101611_21112024_PF_FP_ABST
Abstract
Description
A distributed energy bidirectional prediction optimization scheduling method and system Technical Field
[0001] The present invention relates to the technical field of power grid operation planning, and in particular to a distributed energy bidirectional prediction optimization scheduling method and system. Background Art
[0002] With the rapid development of distributed energy sources such as wind power, photovoltaics, and modern biomass, their share in distribution networks is steadily increasing. Because the generation of new distributed energy sources like wind and photovoltaics is often random, volatile, and intermittent, traditional distribution network dispatching methods are no longer adequate. Therefore, it is urgent to study the safe and economic dispatching of distribution networks in this context of high penetration of new distributed energy sources.
[0003] Optimal scheduling relies on effective predictions of various influencing factors, including power output and load forecasts based on environmental data. Related technologies use meteorological data to predict photovoltaic or wind power, often using recurrent neural networks, long-short-term memory networks, or transform neural networks. Both recurrent neural networks and long-short-term memory networks use a backpropagation algorithm to calculate the loss function for optimization. The longer the sequence, the more difficult it is to find the gradient, meaning convergence becomes more difficult, resulting in poorer performance. Transform neural network structures can completely eliminate the vanishing and exploding gradient problems, but they can also slow down the prediction process for long-term series. Regarding optimal scheduling strategies, some research explores optimal scheduling strategies based on minimizing power generation costs and environmental remediation costs. However, these constraints often rely on the maximum output of distributed generation, potentially making these strategies unfeasible.
[0004] Summary of the Invention
[0005] The present invention provides a distributed energy bidirectional prediction optimization scheduling method and system, which effectively integrates the environment, load and scheduling, predicts the output of distributed energy and then determines the scheduling plan based on the prediction results, thereby improving the accuracy of the scheduling plan.
[0006] In order to solve the above technical problems, an embodiment of the present invention provides a distributed energy bidirectional prediction optimization scheduling method, comprising:
[0007] Obtain historical output data, historical environmental data, historical load data, and historical dispatch strategies within the monitored area. The historical output data includes historical wind power output data and historical photovoltaic power output data.
[0008] Performing orthogonal wavelet transform and reconstruction processing on the historical output data and the historical environmental data respectively to obtain the historical output reconstruction data matrix and the historical environmental reconstruction data matrix, and then extracting and weighted fusion the historical output reconstruction data matrix and the historical environmental reconstruction data matrix through a prediction model based on global features and mutation features to obtain the first output prediction result;
[0009] After solving the reverse distributed energy output prediction problem by using the adaptive prediction model, the historical output data, historical load data and historical dispatch strategy are used to obtain the second output prediction result;
[0010] Performing a composite weighted fusion on the first output prediction result and the second output prediction result to obtain a third output prediction result;
[0011] A distributed energy optimization scheduling model is constructed using the third output prediction result, and the distributed energy optimization scheduling model is solved to obtain a scheduling result, so that the distribution network can schedule energy according to the scheduling result, wherein the scheduling result includes the photovoltaic power generation power value, the wind power generation power value, the energy storage system operating power value and the load shedding power value.
[0012] In the implementation of this embodiment, historical output data, historical environmental data, historical load data, and historical scheduling strategies in the monitored area are obtained, and orthogonal wavelet transform and reconstruction processing are performed on the historical output data and historical environmental data respectively to obtain a historical output reconstruction data matrix, a historical environmental reconstruction data matrix, and a wind speed single-branch reconstruction data matrix. The historical output reconstruction data matrix, the historical environmental reconstruction data matrix, and the wind speed single-branch reconstruction data matrix are subjected to feature extraction and weighted fusion through a prediction model based on global features and mutation features to obtain a first output prediction result; after the historical output data is subjected to reverse distributed energy output prediction and solution through an adaptive prediction model, a second output prediction result is obtained, and a composite weighted fusion is performed on the first output prediction result and the second output prediction result to obtain a third output prediction result. The third output prediction result is used to construct a distributed energy optimization scheduling model, and the scheduling result is obtained after the solution is performed. This method effectively integrates the environment, load, and scheduling, uses two-way prediction to further improve the accuracy of distributed energy output prediction, establishes optimized scheduling on the basis of predicted output, and optimizes the overall scheduling of massive distributed energy.
[0013] As a preferred solution, the historical output data and the historical environmental data are subjected to orthogonal wavelet transform and reconstruction processing respectively to obtain the historical output reconstruction data matrix and the historical environmental reconstruction data matrix, specifically:
[0014] Each data in the historical output data and historical environmental data obtained within the preset time period is normalized to obtain normalized historical output data and normalized historical environmental data, wherein the historical environmental data includes historical ambient light intensity data, historical cut-in wind speed data, and historical cut-out wind speed data, wherein the normalization process is as follows:
[0015] Among them, P WH Represents the historical wind power output data of renewable energy, P PVH Represents historical photovoltaic power generation data, I H Indicates the historical ambient light intensity, V ciH Indicates the historical cut-in wind speed, V coH Indicates historical cut-out wind speed, P WH,min ,P PVH,min ,I H,min , V ciH,min and V coH,min and P WH,max ,P PVH,max ,P LH,max ,I H,max , V ciH,max and V coH,max They are respectively represented by historical wind power output data, historical photovoltaic power output data, historical ambient light intensity, historical cut-in wind speed, historical minimum and maximum cut-out wind speed, and P′ WH , P′ PVH , I′ H , V′ ciH and V′ coH They are respectively represented by the normalized historical wind power output data, historical photovoltaic power output data, historical ambient light intensity, historical cut-in wind speed, and historical cut-out wind speed;
[0016] Each data in the normalized historical output data and the normalized historical environmental data is subjected to positive wavelet transform decomposition according to a preset transformation formula to obtain the real part tree decomposition wavelet coefficient and the imaginary part tree decomposition wavelet coefficient corresponding to each data in the normalized historical output data and the normalized historical environmental data. The preset transformation formula is:
[0017] Where x(t) is the signal to be processed, represents the real part tree decomposition wavelet coefficient, represents the real part tree decomposition scale coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) is the real part of the wavelet function, ψ g(t) is the imaginary wavelet function, J represents the number of layers of wavelet decomposition, and j represents the scale factor;
[0018] According to the real part tree decomposition wavelet coefficient and the imaginary part tree decomposition wavelet coefficient, the wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal corresponding to each data are obtained through the reconstruction formula. The reconstruction formula is:
[0019] Among them, d j (t) represents the reconstructed signal of a certain data wavelet coefficient, c J (t) represents the signal reconstructed by a certain data scaling coefficient, t = 1, 2, ..., M, M is the length of the signal, n is the sequence number of the signal, n = 1, 2, ..., M; represents the real part tree decomposition wavelet coefficient, represents the real part tree decomposition scale coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) represents the real part of the wavelet function, ψ g (t) represents the imaginary wavelet function, J represents the number of layers of wavelet decomposition, and j represents the scale factor;
[0020] The matrix is constructed based on the wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal of each data to obtain the historical output reconstruction data matrix and the historical environment reconstruction data matrix.
[0021] As a preferred solution, the historical output reconstruction data matrix and the historical environment reconstruction data matrix are subjected to feature extraction and weighted fusion through a prediction model based on global features and mutation features to obtain the first output prediction result, specifically:
[0022] The historical output reconstruction data matrix and the historical environment reconstruction data matrix are input into the prediction model based on global features and mutation features. The global trend feature layer is used to extract each data to obtain the first matrix. The local mutation feature layer is used to extract each data to obtain the second matrix. The steady-state feature is used to extract each data to obtain the third matrix. The calculation formula of the steady-state feature is: t =φ0+φ1x t-1 +φ2x t-2 +…+φ p x t-p +ε t
[0023] Among them, x t represents the data at time t, that is, the extracted data, φ i (i=1,…,p) is the autoregressive model coefficient, ε tIndicates that the mean is 0 and the variance is σ 2 White noise, p represents the data length of the current prediction moment value;
[0024] After performing convolution and distillation operations on the first matrix and the second matrix respectively, a new first matrix and a new second matrix are obtained;
[0025] After the new first matrix and the new second matrix are fused to obtain a fourth matrix, the third matrix and the fourth matrix are weightedly fused to obtain a first output prediction result.
[0026] As a preferred solution, after performing convolution and distillation operations on the first matrix and the second matrix respectively, a new first matrix and a new second matrix are obtained, specifically:
[0027] After converting the first matrix and the second matrix into vector matrices in sequence, multiplying the first matrix by a plurality of preset weight matrices respectively to obtain a first vector matrix, a second vector matrix, and a third vector matrix corresponding to the first matrix;
[0028] An approximation evaluation calculation is performed on the first vector matrix according to a preset number of dot products to obtain an approximation evaluation standard for each vector in the first vector matrix, wherein the approximation evaluation standard is:
[0029] Among them, L Q and L K Represented as the first vector matrix and the second vector matrix The length, q i and k j are the i-th and j-th data of the first vector matrix Q1 and the second vector matrix K1, the superscript T is the transpose operation, and d represents the dimension of the matrices Q1, K1, and V1;
[0030] According to the approximation evaluation criteria of each vector in the first vector matrix, the contribution value of each vector to the calculation of the attention value is obtained, the vector with the largest contribution value is selected as the fifth matrix, and the probabilistic sparse attention value is obtained according to the attention value calculation formula based on the fifth matrix, where the probabilistic sparse attention value is:
[0031] in, represents the probabilistic sparse attention value, Q1 represents the first vector matrix, K1 represents the second vector matrix, V1 represents the third vector matrix, d is the dimension of the matrices Q1, K1, and V1, and the superscript T is the transpose operation;
[0032] After performing convolution and distillation operations according to the probabilistic sparse attention value and the preset weight, a new first matrix and a new second matrix are obtained, wherein the convolution and distillation operation process is:
[0033] in,[·] AB Represents the value obtained after an attention module, Conv1d is a one-dimensional convolution, ELU is the activation function of the activation layer, MaxPool is the maximum pooling, is time t The data after the j+1th distillation layer is the new first matrix;
[0034] As a preferred solution, the historical output data, historical load data and historical dispatch strategy are used to perform reverse distributed energy output prediction through an adaptive prediction model to obtain the second output prediction result, which is specifically:
[0035] Expanding the historical scheduling strategy to obtain a new historical scheduling strategy;
[0036] An adaptive prediction model is constructed based on the historical load data, the new historical scheduling strategy, the historical cut-in wind speed, the historical cut-out wind speed and the historical ambient light intensity, and the adaptive prediction model is solved to obtain a second output prediction result, wherein the adaptive prediction model includes a wind adaptive prediction model and a photovoltaic adaptive prediction module, and the second output prediction result includes a wind output prediction result and a photovoltaic output prediction result.
[0037] In this embodiment, the historical scheduling strategy is expanded to obtain a new historical scheduling strategy. An adaptive prediction model is constructed based on the historical load data, the new historical scheduling strategy, the historical cut-in wind speed, the historical cut-out wind speed, and the historical ambient light intensity. This adaptive prediction model is then solved to obtain a second output forecast result. This method considers the scheduling strategy, load, and weather forecasts, as well as different historical scheduling requirements. By effectively integrating the environment, load, and scheduling, more accurate predictions can be achieved.
[0038] As a preferred solution, an adaptive prediction model is constructed based on the historical load data, the new historical scheduling strategy, the historical cut-in wind speed, the historical cut-out wind speed, and the historical ambient light intensity, and the adaptive prediction model is solved to obtain a second output prediction result, specifically:
[0039] A wind adaptive prediction model is constructed based on the historical load data, the new historical dispatch strategy, the historical cut-in wind speed, and the historical cut-out wind speed, wherein the wind adaptive prediction model is: WH =β Wi X Wi +β W
[0040] Among them, PWH represents the wind power output prediction result, β Wi (i=1,2,3,4) is the regression coefficient between wind power output and related factors, β W is the overall error of the wind output model, X Wi represents the wind power output prediction factor matrix, X Wi =[P LH ,D H ,V ciH ,V coH ], where P LH Indicates historical load data, D H Represents the new historical scheduling strategy, V ciH Indicates the historical cut-in wind speed and V coH Historical cut-out wind speed;
[0041] Solve P using the least squares method WH =β Wi X Wi +β W Regression coefficient, get the initial value of regression coefficient Further introduction of weight Then the weight vector It can be expressed as:
[0042] renew x j It's X Wi Variables in
[0043] For all regularization factors λ n , solve
[0044] calculate Will Substitute P WH =β Wi X Wi +β W , get the wind power output prediction result;
[0045] For photovoltaic output, an adaptive prediction model is established based on historical load, new historical scheduling strategy, and historical ambient light intensity: PVH =β PVi X PVi +β PV
[0046] Among them, P PVH represents the photovoltaic output regression model, β PVi (i=1,2,3) is the regression coefficient between photovoltaic output and related factors, β PV is the overall error of the photovoltaic output model, X PVirepresents the photovoltaic output prediction factor matrix, X PVi =[P LH ,D H ,I H ], where P LH Indicates historical load data, D H Represents the new historical scheduling strategy and I H Indicates the historical ambient light intensity;
[0047] Solve P using the least squares method PVH =β PVi X PVi +β PV The regression coefficient is obtained by Further introduction of weight Then the weight vector It can be expressed as:
[0048] renew x j It's X PVi Variables in
[0049] For all regularization factors λ n , solve
[0050] calculate Will Substitute P PVH =β PVi X PVi +β PV , and obtain the prediction result of photovoltaic output.
[0051] As a preferred solution, a composite weighted fusion is performed on the first output prediction result and the second output prediction result to obtain a third output prediction result, specifically:
[0052] The first output prediction result and the second output prediction result are fused by weighted average method to obtain the third output prediction result. The specific calculation formula is: WTP =θP WP1 +(1-θ)P WP2
[0053] Among them, P WTP Indicates the third output prediction result, P WP1 Indicates the first output prediction result, P WP2 represents the second output prediction result, and θ represents the power coefficient predicted by the forward prediction method.
[0054] As a preferred solution, the third output prediction result is used to build a distributed energy optimization scheduling model. The distributed energy optimization scheduling model is used to optimize the historical output data and energy storage equipment operation data to obtain the scheduling results, specifically:
[0055] The power generation cost is obtained by constructing the power generation cost according to the third prediction result and the power generation cost coefficient, wherein the power generation cost is the photovoltaic power generation cost or the wind power generation cost;
[0056] The operation and management cost of the energy storage equipment is obtained based on the charging operation cost coefficient, discharging operation cost coefficient, charging efficiency, discharging efficiency, charging power and discharging power;
[0057] The environmental governance cost of the energy storage equipment is obtained by constructing the cost coefficient of the energy storage equipment operating environment, the charging power of the energy storage equipment, and the discharging power of the energy storage equipment. The load shedding cost is obtained by constructing the load shedding penalty coefficient and the load shedding power.
[0058] The objective function and constraints of the distributed energy optimization scheduling model are constructed based on the power generation cost, energy storage equipment operation and management cost, energy storage equipment environmental management cost and load shedding cost, wherein the objective function is: min(C PV +C WT +C SV +C SVE +C loss )
[0059] Among them, C PV is the photovoltaic power generation cost, C WT is the cost of wind power generation, C SV is the operation and management cost of energy storage equipment, C SVE is the environmental management cost of energy storage equipment, C loss is the load shedding cost;
[0060] The alternating direction multiplier method is used to solve the objective function and constraints to obtain the scheduling results, which include the photovoltaic power generation value, wind power generation value, energy storage system operating power value and load shedding power value.
[0061] In this embodiment, a model is constructed based on the third prediction result and the power generation cost coefficient to obtain the power generation cost, which is either the photovoltaic power generation cost or the wind power generation cost. The model is constructed based on the charging operation cost coefficient, the discharging operation cost coefficient, the charging efficiency, the discharging efficiency, the charging power, and the discharging power to obtain the energy storage device operation and management cost. The model is constructed based on the energy storage device operation environment management cost coefficient, the energy storage device charging power, and the energy storage device discharging power to obtain the energy storage device environmental management cost. The model is constructed based on the load shedding penalty coefficient and the load shedding power to obtain the load shedding cost. The objective function and constraints of the distributed energy optimization scheduling model are constructed based on the power generation cost, the energy storage device operation and management cost, the energy storage device environmental management cost, and the load shedding cost. The alternating direction multiplier method is used to solve the objective function and constraints to obtain a scheduling result, which includes the photovoltaic power generation power value, the wind power generation power value, the energy storage system operation power value, and the load shedding power value. By utilizing the optimization scheduling model based on the prediction results and solving and outputting the optimization scheduling result, the accuracy of the overall distributed energy scheduling plan is improved.
[0062] As a preferred solution, in order to solve the same technical problem, an embodiment of the present invention further provides a distributed energy bidirectional prediction optimization scheduling system, comprising:
[0063] An acquisition module is used to obtain historical output data, historical environmental data, historical load data, and historical scheduling strategies in the monitored area, wherein the historical output data includes historical wind power output data and historical photovoltaic power output data;
[0064] The first prediction module is used to perform orthogonal wavelet transform and reconstruction processing on the historical output data and the historical environmental data, respectively, to obtain a historical output reconstruction data matrix and a historical environmental reconstruction data matrix, and perform feature extraction and weighted fusion on the historical output reconstruction data matrix and the historical environmental reconstruction data matrix through a prediction model based on global features and mutation features to obtain a first output prediction result;
[0065] The second prediction module is used to perform reverse distributed energy output prediction and solve the historical output data, historical load data and historical dispatch strategy through the adaptive prediction model to obtain a second output prediction result;
[0066] A third prediction module is used to perform a composite weighted fusion on the first output prediction result and the second output prediction result to obtain a third output prediction result;
[0067] The scheduling result module is used to use the third output prediction result to build a distributed energy optimization scheduling model, and solve the distributed energy optimization scheduling model to obtain the scheduling result, so that the distribution network can schedule energy according to the scheduling result. The scheduling result includes the photovoltaic power generation power value, the wind power generation power value, the energy storage system operating power value and the load shedding power value.
[0068] As a preferred solution, the first prediction module includes a normalization unit, a decomposition unit, a reconstruction unit and a matrix unit.
[0069] The normalization unit is used to normalize each data in the historical output data and historical environmental data obtained within a preset time period to obtain normalized historical output data and normalized historical environmental data, wherein the historical environmental data includes historical ambient light intensity data, historical cut-in wind speed data, and historical cut-out wind speed data. The normalization process is as follows:
[0070] Among them, P WH Represents the historical wind power output data of renewable energy, P PVH Represents historical photovoltaic power generation data, I H Indicates the historical ambient light intensity, V ciH Indicates the historical cut-in wind speed, V coH Indicates historical cut-out wind speed, P WH,min ,P PVH,min ,I H,min , V ciH,min and V coH,min and P WH,max ,P PVH,max ,P LH,max ,I H,max , V ciH,max and V coH,max They are respectively represented by historical wind power output data, historical photovoltaic power output data, historical ambient light intensity, historical cut-in wind speed, historical minimum and maximum cut-out wind speed, and P′ WH , P′ PVH 、I' H , V′ ciH and V′ coH They are respectively represented by the normalized historical wind power output data, historical photovoltaic power output data, historical ambient light intensity, historical cut-in wind speed, and historical cut-out wind speed;
[0071] The decomposition unit is used to perform positive wavelet transform decomposition on each data in the normalized historical output data and the normalized historical environmental data according to a preset transformation formula, to obtain the real part tree decomposition wavelet coefficient and the imaginary part tree decomposition wavelet coefficient corresponding to each data in the normalized historical output data and the normalized historical environmental data, wherein the preset transformation formula is:
[0072] Where x(t) is the signal to be processed, represents the real part tree decomposition wavelet coefficient, represents the real part tree decomposition scale coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) is the real part of the wavelet function, ψ g (t) is the imaginary wavelet function, J represents the number of layers of wavelet decomposition, and j represents the scale factor;
[0073] The reconstruction unit is used to calculate the wavelet coefficient reconstruction signal and scale coefficient reconstruction signal corresponding to each data according to the real part tree decomposition wavelet coefficient and the imaginary part tree decomposition wavelet coefficient through the reconstruction formula, wherein the reconstruction formula is:
[0074] Among them, d j (t) represents the reconstructed signal of a certain data wavelet coefficient, c J (t) represents the signal reconstructed by a certain data scaling coefficient, t = 1, 2, ..., M, M is the length of the signal, n is the sequence number of the signal, n = 1, 2, ..., M; represents the real part tree decomposition wavelet coefficient, represents the real part tree decomposition scale coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) represents the real part of the wavelet function, ψ g (t) represents the imaginary wavelet function, J represents the number of layers of wavelet decomposition, and j represents the scale factor;
[0075] The matrix unit is used to construct a matrix based on the wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal of each data to obtain the historical output reconstruction data matrix and the historical environment reconstruction data matrix.
[0076] The method obtains historical output data, historical environmental data, historical load data, and historical dispatch strategies for the monitored area, performs orthogonal wavelet transform and reconstruction on the historical output data and historical environmental data, and obtains a historical output reconstruction data matrix, a historical environmental reconstruction data matrix, and a wind speed single-branch reconstruction data matrix. These matrices are then subjected to feature extraction and weighted fusion using a prediction model based on global and mutation features to obtain a first output forecast result. The historical output data is then used to perform reverse distributed energy output forecasting using an adaptive prediction model to obtain a second output forecast result. The first and second output forecast results are then compositely weightedly fused to obtain a third output forecast result. The third output forecast result is used to construct a distributed energy optimization dispatch model and solve it to obtain a dispatch result. This method effectively integrates the environment, load, and dispatch, uses bidirectional prediction to further improve the accuracy of distributed energy output forecasting, and establishes optimized dispatching based on predicted output, thereby optimizing the overall dispatching of massive distributed energy resources. BRIEF DESCRIPTION OF THE DRAWINGS
[0077] FIG1 is a flow chart of an embodiment of a method for bidirectional prediction and optimization scheduling of distributed energy resources provided by the present invention;
[0078] FIG2 is a schematic diagram of a data calculation process for an embodiment of a distributed energy bidirectional prediction optimization scheduling method provided by the present invention;
[0079] FIG3 is a schematic diagram of a first prediction process of an embodiment of a method for bidirectional prediction and optimization scheduling of distributed energy resources provided by the present invention;
[0080] FIG4 is a schematic diagram of a second prediction process of an embodiment of the distributed energy bidirectional prediction optimization scheduling method provided by the present invention;
[0081] FIG5 is a schematic diagram of the system structure of another embodiment of the distributed energy bidirectional prediction optimization scheduling method provided by the present invention; DETAILED DESCRIPTION
[0082] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.
[0083] Example 1
[0084] Referring to FIG. 1 , a distributed energy bidirectional prediction optimization scheduling method provided by an embodiment of the present invention includes steps 101 to 104 , each of which is specifically as follows:
[0085] Step 101: Acquire historical output data, historical environmental data, historical load data, and historical dispatch strategies in the area to be monitored, wherein the historical output data includes historical wind power output data and historical photovoltaic power output data.
[0086] In this embodiment, historical renewable energy output data, historical environmental data, and historical load data within a preset time period are obtained in the monitoring area, and a historical dispatch strategy is obtained. The historical output data includes historical wind power generation output data and historical photovoltaic power generation output data, and the historical environmental data includes historical ambient light intensity, historical cut-in wind speed, and historical cut-out wind speed data.
[0087] It should be noted that the time period and time interval for acquiring data can be determined according to actual needs.
[0088] As an example of this embodiment, historical wind power output data, historical photovoltaic power output data, historical load data, historical ambient light intensity, historical cut-in wind speed and historical cut-out wind speed in the monitoring area within 30 days before the current date at intervals of 15 minutes can be obtained.
[0089] Step 102: Perform orthogonal wavelet transform and reconstruction processing on the historical output data and the historical environmental data respectively to obtain a historical output reconstruction data matrix and a historical environmental reconstruction data matrix, and perform feature extraction and weighted fusion on the historical output reconstruction data matrix and the historical environmental reconstruction data matrix through a prediction model based on global features and mutation features to obtain a first output prediction result.
[0090] Optionally, the historical output data and the historical environment data are subjected to orthogonal wavelet transform and reconstruction processing respectively to obtain a historical output reconstruction data matrix and a historical environment reconstruction data matrix, specifically:
[0091] Each data in the historical output data and historical environmental data obtained within the preset time period is normalized to obtain normalized historical output data and normalized historical environmental data, wherein the historical environmental data includes historical ambient light intensity data, historical cut-in wind speed data, and historical cut-out wind speed data. The normalization process is as follows:
[0092] Among them, P WH Represents the historical wind power output data of renewable energy, P PVH Represents historical photovoltaic power generation data, P LH Indicates historical load data, IH Indicates the historical ambient light intensity, V ciH Indicates the historical cut-in wind speed, V coH Indicates historical cut-out wind speed, P WH,min ,P PVH,min ,I H,min , V ciH,min and V coH,min and P WH,max ,P PVH,max ,P LH,max ,I H,max , V ciH,max and V coH,max They are respectively represented by historical wind power output data, historical photovoltaic power output data, historical ambient light intensity, historical cut-in wind speed, historical minimum and maximum cut-out wind speed, and P′ WH , P′ PVH 、I' H , V′ ciH and V′ coH They are respectively represented by the normalized historical wind power output data, historical photovoltaic power output data, historical ambient light intensity, historical cut-in wind speed, and historical cut-out wind speed;
[0093] Each data in the normalized historical output data and the normalized historical environmental data is decomposed by positive wavelet transform according to a preset transformation formula to obtain the real part tree decomposition wavelet coefficient and the imaginary part tree decomposition wavelet coefficient corresponding to each data in the normalized historical output data and the normalized historical environmental data. The preset transformation formula is:
[0094] Where x(t) is the signal to be processed, represents the real part tree decomposition wavelet coefficient, represents the real part tree decomposition scale coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) is the real part of the wavelet function, ψ g (t) is the imaginary wavelet function, J represents the number of layers of wavelet decomposition, and j represents the scale factor;
[0095] According to the real part tree decomposition wavelet coefficient and the imaginary part tree decomposition wavelet coefficient, the wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal corresponding to each data are obtained through the reconstruction formula. The reconstruction formula is:
[0096] Among them, d j (t) represents the reconstructed signal of a certain data wavelet coefficient, cJ (t) represents the reconstructed signal of a certain data scale coefficient, t = 1, 2, ..., M, M is the length of the reconstructed signal, n is the sequence number of the reconstructed signal, n = 1, 2, ..., M; represents the real part tree decomposition wavelet coefficient, represents the real part tree decomposition scale coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) represents the real part of the wavelet function, ψ g (t) represents the imaginary wavelet function, J represents the number of layers of wavelet decomposition, and j represents the scale factor;
[0097] The matrix is constructed based on the wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal of each data to obtain the historical output reconstruction data matrix and the historical environment reconstruction data matrix.
[0098] In this embodiment, as shown in FIG2 , in order to eliminate the influence of different data dimensions, the acquired historical output data and historical environmental data are normalized and pre-processed to obtain normalized historical output data and normalized historical environmental data. The historical environmental data includes historical ambient light intensity data and historical cut-in wind speed data. The normalization process of the historical cut-out wind speed data is as follows:
[0099] Among them, P WH Represents the historical wind power output data of renewable energy, P PVH Represents historical photovoltaic power generation data, P LH Indicates historical load data, I H Indicates the historical ambient light intensity, V ciH Indicates the historical cut-in wind speed, V coH Indicates historical cut-out wind speed, P WH,min ,P PVH,min ,P LH,min ,I H,min , V ciH,min and V coH,min and P WH,max ,P PVH,max ,P LH,max ,I H,max , V ciH,max and V coH,max They are respectively represented by historical wind power output data, historical photovoltaic power output data, historical load data, historical ambient light intensity, historical cut-in wind speed, historical minimum and maximum cut-out wind speed, and P′ WH, P′ PVH , P′ LH 、I' H , V′ ciH and V′ coH They are respectively represented by the normalized historical wind power output data, historical photovoltaic power output data, historical load data, historical ambient light intensity, historical cut-in wind speed, and historical cut-out wind speed.
[0100] After obtaining the normalized historical wind power output data, historical photovoltaic power output data, historical load data, historical ambient light intensity, historical cut-in wind speed, and historical cut-out wind speed data, orthogonal wavelet transform and reconstruction are performed on the historical wind power output data, historical photovoltaic power output data, historical load data, historical ambient light intensity, historical cut-in wind speed, and historical cut-out wind speed data. The specific method is as follows:
[0101] All data are decomposed according to the following formula to obtain the real part tree decomposition wavelet coefficient and the imaginary part tree decomposition wavelet coefficient of each data. The decomposition formula is:
[0102] Where x(t) is the signal to be processed, represents the real part tree decomposition wavelet coefficient, represents the real part tree decomposition scale coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) is the real part of the wavelet function, ψ g (t) is the imaginary wavelet function, J represents the number of wavelet decomposition layers, and j represents the scale factor.
[0103] Then, the wavelet coefficients decomposed by the real part tree and the wavelet coefficients decomposed by the imaginary part tree are calculated through the reconstruction formula to obtain the wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal corresponding to each data. Then, a matrix is constructed based on the wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal of each data to obtain the historical output reconstruction data matrix and the historical environment reconstruction data matrix. The reconstruction formula is:
[0104] Among them, d j (t) represents the reconstructed signal of a certain data wavelet coefficient, c J (t) represents the signal reconstructed by a certain data scaling coefficient, t = 1, 2, ..., M, M is the length of the signal, n is the sequence number of the signal, n = 1, 2, ..., M; represents the real part tree decomposition wavelet coefficient, represents the real part tree decomposition scale coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) represents the real part of the wavelet function, ψ g (t) represents the imaginary wavelet function, J represents the number of wavelet decomposition layers, and j represents the scale factor.
[0105] As an example of this embodiment, select J=4, and 5 sets of data can be obtained after single-branch reconstruction, as described in the following formula: RCWT =[c I4 ,d I4 ,d I3 ,d I2 ,d I1 ] V ciRCWT =[c ci4 ,d ci4 ,d ci3 ,d ci2 ,d ci1 ] V coRCWT =[c co4 ,d co4 ,d co3 ,d co2 ,d co1 ]
[0106] Among them, I RCWT Represented as the single-branch reconstruction data matrix of historical environmental data, c I4 Represented as the reconstructed signal of the illumination data scale coefficient, d I4 ,d I3 ,d I2 ,d I1 Represented as the reconstruction signal of the wavelet coefficient of the illumination data, V ciRCWT It is represented by the reconstructed data matrix of the single wind speed in the historical environmental data, c ci4 It is expressed as the reconstructed signal of the cut-in wind speed data scale coefficient, d ci4 ,d ci3 ,d ci2 ,d ci1 It is represented by the cut-in wind speed data wavelet coefficient reconstruction signal, V coRCWT It is represented by the reconstructed data matrix of the wind speed single branch cut out from the historical environmental data, c co4 It is expressed as the reconstructed signal of the scale coefficient of the cut-out wind speed data, d co4 ,d co3 ,d co2 ,d co1 It represents the reconstructed signal of the wavelet coefficients of the cut-out wind speed data.
[0107] Optionally, the historical output reconstruction data matrix and the historical environment reconstruction data matrix are subjected to feature extraction and weighted fusion through a prediction model based on global features and mutation features to obtain a first output prediction result, specifically:
[0108] The historical output reconstruction data matrix and the historical environment reconstruction data matrix are input into the prediction model based on global features and mutation features. The global trend feature layer is used to extract each data to obtain the first matrix. The local mutation feature layer is used to extract each data to obtain the second matrix. The steady-state feature is used to extract each data to obtain the third matrix. The calculation formula of the steady-state feature is: t =φ0+φ1x t-1 +φ2x t-2 +…+φ p x t-p +ε t
[0109] Among them, x t represents the data at time t, that is, the extracted data, φ i (i=1,…,p) is the autoregressive model coefficient, ε t Indicates that the mean is 0 and the variance is σ 2 White noise, p represents the data length of the current prediction moment value;
[0110] After performing convolution and distillation operations on the first matrix and the second matrix respectively, a new first matrix and a new second matrix are obtained;
[0111] After the new first matrix and the new second matrix are fused to obtain a fourth matrix, the third matrix and the fourth matrix are weightedly fused to obtain a first output prediction result.
[0112] In this embodiment, as shown in Figure 3, the historical output reconstruction data matrix and the historical environment reconstruction data matrix are input into a prediction model based on global features and mutation features. This model is a prediction network based on global trends and local mutations, primarily comprising a global trend feature acquisition branch, a local mutation feature acquisition branch, and a steady-state feature acquisition branch. The global trend feature acquisition branch is used to extract each data point to obtain a first matrix. The local mutation feature layer is used to extract each data point to obtain a second matrix. The steady-state feature layer is used to extract each data point to obtain a third matrix. After convolution and distillation operations are performed on the first and second matrices, respectively, through the embedding layer and the distillation layer, new first and second matrices are obtained. The new first and second matrices are then subjected to feature fusion in a fully connected layer to obtain a fourth matrix. The third and fourth matrices are then weightedly fused to produce the first output prediction result.
[0113] Optionally, after performing convolution and distillation operations on the first matrix and the second matrix respectively, a new first matrix and a new second matrix are obtained, specifically:
[0114] After converting the first matrix and the second matrix into vector matrices in sequence, multiplying the first matrix by a plurality of preset weight matrices respectively to obtain a first vector matrix, a second vector matrix, and a third vector matrix corresponding to the first matrix;
[0115] An approximation evaluation calculation is performed on the first vector matrix according to a preset number of dot products to obtain an approximation evaluation standard for each vector in the first vector matrix, wherein the approximation evaluation standard is:
[0116] Among them, L Q and L K Represented as the first vector matrix and the second vector matrix The length, q i and k j are the i-th and j-th data of the first vector matrix Q1 and the second vector matrix K1, the superscript T is the transpose operation, and d represents the dimension of the matrices Q1, K1, and V1;
[0117] According to the approximation evaluation criteria of each vector in the first vector matrix, the contribution value of each vector to the calculation of the attention value is obtained, the vector with the largest contribution value is selected as the fifth matrix, and the probabilistic sparse attention value is obtained according to the attention value calculation formula based on the fifth matrix, wherein the probabilistic sparse attention value is:
[0118] in, represents the probabilistic sparse attention value, Q1 represents the first vector matrix, K1 represents the second vector matrix, V1 represents the third vector matrix, d is the dimension of the matrices Q1, K1, and V1, and the superscript T is the transpose operation;
[0119] After performing convolution and distillation operations according to the probabilistic sparse attention value and the preset weight, a new first matrix and a new second matrix are obtained, wherein the convolution and distillation operation process is:
[0120] in,[·] AB Represents the value obtained after an attention module, Conv1d is a one-dimensional convolution, ELU is the activation function of the activation layer, MaxPool is the maximum pooling, is time t The data after the j+1th distillation layer is the new first matrix.
[0121] In this embodiment, in the global trend feature acquisition branch, after the historical output reconstruction data matrix and the historical environment reconstruction data matrix are input into the global trend feature acquisition branch, the data is feature extracted to obtain the first matrix. It should be noted that if the input data has D feature variables, the filter scale is N, and the number of filters is m g , then the dimension of the first matrix is D×m g ; After extracting features from each data using the local mutation feature layer, the second matrix is obtained. If the input data has D feature variables, the filter scale is L, the length of L is less than N, and the number of filters is m l , after obtaining the branch through local mutation features, the second matrix dimension is D×m l .
[0122] The vector matrix after the first matrix is transformed by the embedding layer is randomly generated into three weight matrices W q 、W k 、W v , multiply the first matrix by the three weight matrices respectively to obtain the first vector matrix, the second vector matrix and the third vector matrix corresponding to the first matrix, Q1=W q G1 K1=W k G1 V1=W v G1
[0123] Among them, Q1 represents the first vector matrix, K1 represents the second vector matrix, V1 represents the third vector matrix, W q represents the first weight matrix, W k represents the second weight matrix, W v represents the third matrix weight, and G1 represents the first matrix.
[0124] Randomly select L Q Ln K The approximate evaluation standard formula for the dot product pair calculates the approximate evaluation standard of the sparsity of each vector in the first vector matrix. The approximate evaluation standard formula is:
[0125] Among them, L Q and L K Represented as the first vector matrix and the second vector matrix The length, q i and k j are the i-th and j-th data of the first vector matrix Q1 and the second vector matrix K1, the superscript T is the transpose operation, and d represents the dimension of the matrices Q1, K1, and V1.
[0126] According to the approximation evaluation criteria of each vector in the first vector matrix, the part of the first vector matrix that contributes most to the attention value is selected as the fifth matrix, and the probabilistic sparse attention value is obtained according to the attention value calculation formula based on the fifth matrix. The attention value calculation formula is:
[0127] in, represents the probabilistic sparse attention value, Q1 represents the first vector matrix, K1 represents the second vector matrix, V1 represents the third vector matrix, d is the dimension of the matrices Q1, K1, and V1, and the superscript T is the transpose operation.
[0128] The weights of the main features are set higher. After the distillation layer from layer j to layer j+1, the new first matrix is obtained by the following operations:
[0129] in,[·] AB Represents the value obtained after an attention module, Conv1d is a one-dimensional convolution, ELU is the activation function of the activation layer, MaxPool is the maximum pooling, is time t The data after the j+1th distillation layer is the new first matrix.
[0130] Then, the vector matrix after the second matrix is transformed by the embedding layer is randomly generated into three weight matrices M q 、M k 、M v, Multiply the second matrix by the three weight matrices respectively to obtain the first vector matrix, the second vector matrix and the third vector matrix corresponding to the second matrix, Q2=M q G2 K2=M k G2 V2=M v G2
[0131] Among them, Q2 represents the first vector matrix, K2 represents the second vector matrix, V2 represents the third vector matrix, M q represents the first weight matrix, M k represents the second weight matrix, M v Represents the third matrix weight.
[0132] Randomly select L Q Ln K The approximate evaluation standard formula for the dot product pair calculates the approximate evaluation standard of the sparsity of each vector in the first vector matrix. The approximate evaluation standard formula is:
[0133] Among them, L Q and L K Represented as the first vector matrix and the second vector matrix The length, q i and k j are the i-th and j-th data of the first vector matrix and the second vector matrix, the superscript T is the transpose operation, and d is the dimension of the matrices Q2, K2, and V2.
[0134] According to the approximation evaluation criteria of each vector in the first vector matrix, the part of the first vector matrix that contributes the most to the attention value is selected as the fifth matrix, and the probabilistic sparse attention value is obtained according to the attention value calculation formula based on the fifth matrix. The attention value calculation formula is:
[0135] in, represents the probabilistic sparse attention value, Q2 represents the first vector matrix, K2 represents the second vector matrix, V2 represents the third vector matrix, d is the dimension of the matrices Q2, K2, and V2, and the superscript T is the transpose operation.
[0136] The weights of the main features are set higher. After the distillation layer j to j+1 layer, the new second matrix is obtained by the following operations:
[0137] in,[·] AB Represents the value obtained after an attention module, Conv1d is a one-dimensional convolution, ELU is the activation function of the activation layer, MaxPool is the maximum pooling, is time t Data after the j+1th distillation layer.
[0138] Use the steady-state characteristics to extract each data and obtain the third matrix. The calculation formula is: t =φ0+φ1x t-1 +φ2x t-2 +…+φ p x t-p +ε t
[0139] Among them, x t represents the data value at time t, φ i (i=1,…,p) is the autoregressive model coefficient, ε t Indicates that the mean is 0 and the variance is σ 2 white noise, and p represents the data length of the current prediction moment value.
[0140] The new first matrix obtained by the global trend feature acquisition branch and the new second matrix obtained by the local mutation feature acquisition branch are input into the fully connected layer to obtain the feature matrix F t Then, the steady-state characteristic x t With the feature matrix Ft Perform weighted fusion to obtain the final wind power output or photovoltaic power output prediction result.
[0141] Step 103: After performing reverse distributed energy output prediction on the historical output data, historical load data, and historical dispatch strategies through an adaptive prediction model, a second output prediction result is obtained.
[0142] Optionally, after performing reverse distributed energy output prediction on historical output data, historical load data, and historical dispatch strategies through an adaptive prediction model, a second output prediction result is obtained, specifically:
[0143] Expand the historical scheduling strategy to obtain a new historical scheduling strategy;
[0144] An adaptive prediction model is constructed based on historical load data, a new historical scheduling strategy, historical cut-in wind speed, historical cut-out wind speed and historical ambient light intensity, and the adaptive prediction model is solved to obtain a second output prediction result, wherein the adaptive prediction model includes a wind adaptive prediction model and a photovoltaic adaptive prediction module, and the second output prediction result includes a wind output prediction result and a photovoltaic output prediction result.
[0145] Optionally, an adaptive prediction model is constructed based on historical load data, a new historical dispatch strategy, historical cut-in wind speed, historical cut-out wind speed, and historical ambient light intensity, and the adaptive prediction model is solved to obtain a second output prediction result, specifically:
[0146] A wind adaptive prediction model is constructed based on historical load data, new historical dispatching strategies, historical cut-in wind speeds, and historical cut-out wind speeds. The wind adaptive prediction model is: P WH =β Wi X Wi +β W
[0147] Among them, P WH represents the wind power output prediction result, β Wi (i=1,2,3,4) is the regression coefficient between wind power output and related factors, β W is the overall error of the wind output model, X Wi represents the wind power output prediction factor matrix, X Wi =[P LH ,D H ,V ciH ,V coH ], where P LH Indicates historical load data, D H Represents the new historical scheduling strategy, V ciH Indicates the historical cut-in wind speed and V coHHistorical cut-out wind speed;
[0148] Solve P using the least squares method WH =β Wi X Wi +β W Regression coefficient, get the initial value of regression coefficient Further introduction of weight Then the weight vector It can be expressed as:
[0149] renew x j It's X Wi Variables in
[0150] For all regularization factors λ n , solve
[0151] calculate Will Substitute P WH =β Wi X Wi +β W , get the wind power output prediction result;
[0152] For photovoltaic output, an adaptive prediction model is established based on historical load, new historical scheduling strategy, and historical ambient light intensity: PVH =β PVi X PVi +β PV
[0153] Among them, P PVH represents the photovoltaic output regression model, β PVi (i=1,2,3) is the regression coefficient between photovoltaic output and related factors, β PV is the overall error of the photovoltaic output model, X PVi represents the photovoltaic output prediction factor matrix, X PVi =[P LH ,D H ,I H ], where P LH Indicates historical load data, D H Represents the new historical scheduling strategy and I H Indicates the historical ambient light intensity;
[0154] Solve P using the least squares method PVH =β PVi X PVi +β PV The regression coefficient is obtained by Further introduction of weight Then the weight vector It can be expressed as:
[0155] renew x j It's X PVi Variables in
[0156] For all regularization factors λ n , solve
[0157] calculate Will Substitute P PVH =β PVi X PVi +β PV , and obtain the photovoltaic output prediction result.
[0158] In this embodiment, as shown in Figure 4, the historical scheduling strategy of distributed energy is first expanded to have the same dimension as other historical data, and then the historical output data, historical load data and historical environmental data of distributed energy are preprocessed to establish a linear regression model of output and load, scheduling strategy and environment. The adaptive prediction model includes a wind adaptive prediction model and a photovoltaic adaptive prediction model.
[0159] For wind power output, an adaptive prediction model is established based on historical load, new historical dispatch strategy, historical cut-in wind speed and historical cut-out wind speed: WH =β Wi X Wi +β W
[0160] Among them, P WH represents the wind power output prediction result, β Wi (i=1,2,3,4) is the regression coefficient between wind power output and related factors, β W is the overall error of the wind output model, X Wi represents the wind power output prediction factor matrix, X Wi =[P LH ,D H ,V ciH ,V coH ], where P LH Indicates historical load data, D H Represents the new historical scheduling strategy, V ciH Indicates the historical cut-in wind speed and V coH Historical cut-out wind speed;
[0161] First, the least squares method is used to solve the regression coefficient of the wind adaptive prediction model to obtain the initial value of the regression coefficient Further introduction of weight Then the weight vector It can be expressed as:
[0162] Next, update x j It's X Wi Variables in
[0163] Furthermore, for all regularization factors λ n , solve
[0164] Further calculation:
[0165] Will Substitute P WH =β Wi X Wi +β W , that is, to obtain the prediction of wind power output;
[0166] For photovoltaic output, an adaptive prediction model is established based on historical load, new historical scheduling strategy, and historical ambient light intensity: PVH =β PVi X PVi +β PV
[0167] Among them, P PVH represents the photovoltaic output regression model, β PVi (i=1,2,3) is the regression coefficient between photovoltaic output and related factors, β PV is the overall error of the photovoltaic output model, X PVi represents the photovoltaic output prediction factor matrix, X PVi =[P LH ,D H ,I H ], where P LH Indicates historical load data, D H Represents the new historical scheduling strategy and I H Indicates the historical ambient light intensity;
[0168] First, the least squares method is used to solve the regression coefficient of the photovoltaic adaptive prediction model to obtain the initial value of the regression coefficient Further introduction of weight Then the weight vector It can be expressed as:
[0169] Next, update x j It's X PVi Variables in
[0170] Furthermore, for all regularization factors λ n , solve
[0171] Further calculation Will Substitute P PVH =β PVi X PVi +β PV , that is, we can get the prediction of photovoltaic output.
[0172] Step 104: Perform composite weighted fusion on the first output prediction result and the second output prediction result to obtain a third output prediction result.
[0173] Optionally, a composite weighted fusion is performed on the first output prediction result and the second output prediction result to obtain a third output prediction result, specifically:
[0174] The first output prediction result and the second output prediction result are fused by weighted average method to obtain the third output prediction result. The specific calculation formula is: HP =θP HP1 +(1-θ)P HP2
[0175] Among them, P HP Indicates the third output prediction result, P HP1 Indicates the first output prediction result, P HP2 represents the second output prediction result, and θ represents the power coefficient predicted by the forward prediction method.
[0176] In this embodiment, the first prediction result and the second prediction result are combined by weighted average method to perform output prediction fusion, and the wind power output prediction result is: WTP =θ1P WP1 +(1-θ1)P WP2
[0177] The photovoltaic output prediction result is: P PVP =θ2P PVP1 +(1-θ2)P PVP2
[0178] Among them, P WTP Represents the predicted value of wind power composite output, P WP1 Represents the first wind force forecast result, P WP2 Represents the wind power output forecast result, P PVP Represents the predicted value of photovoltaic composite output, PPVP1 Represents the first photovoltaic output prediction result, P PVP2 represents the photovoltaic output prediction result, θ1 is the coefficient of wind power output predicted by the forward prediction method, and θ2 is the coefficient of photovoltaic power output predicted by the forward prediction method.
[0179] Step 105: Utilize the third output prediction result to construct a distributed energy optimization scheduling model, and use the distributed energy optimization scheduling model to optimize and schedule the historical output data and the energy storage device operating data to obtain a scheduling result, so that the distribution network can schedule energy according to the scheduling result, wherein the scheduling result includes the photovoltaic power generation value, the wind power generation power value, the energy storage system operating power value, and the load shedding power value.
[0180] Optionally, constructing a power generation cost based on the third prediction result and the power generation cost coefficient to obtain the power generation cost, wherein the power generation cost is the photovoltaic power generation cost or the wind power generation cost;
[0181] The operation and management cost of the energy storage equipment is obtained based on the charging operation cost coefficient, discharging operation cost coefficient, charging efficiency, discharging efficiency, charging power and discharging power;
[0182] The environmental governance cost of the energy storage equipment is obtained by constructing the cost coefficient of the energy storage equipment operating environment, the charging power of the energy storage equipment, and the discharging power of the energy storage equipment. The load shedding cost is obtained by constructing the load shedding penalty coefficient and the load shedding power.
[0183] The objective function and constraints of the distributed energy optimization scheduling model are constructed based on the power generation cost, energy storage equipment operation and management cost, energy storage equipment environmental management cost and load shedding cost. The objective function is: min(C PV +C WT +C SV +C SVE +C loss )
[0184] Among them, C PV is the photovoltaic power generation cost, C WT is the cost of wind power generation, C SV is the operation and management cost of energy storage equipment, C SVE is the environmental management cost of energy storage equipment, C loss is the load shedding cost;
[0185] The alternating direction multiplier method is used to solve the objective function and constraints to obtain the scheduling results, which include the photovoltaic power generation value, wind power generation value, energy storage system operating power value and load shedding power value.
[0186] In this embodiment, the third prediction result, that is, the third predicted power, is used to calculate the power generation cost:
[0187] The cost of photovoltaic power generation is: C PV =k PV P PVT
[0188] Among them, k PV is the photovoltaic power generation cost coefficient, P PVT is the photovoltaic power generation,
[0189] Constraints: P PVmin ≤P PVT ≤P PVP ;
[0190] The cost of wind power generation is: C WT =k WT P WTT
[0191] Among them, k WT is the wind power generation cost coefficient, P WTT is the amount of wind power generated;
[0192] The constraints are: P WTmin ≤P WTT ≤P WTP ;
[0193] The calculation of energy storage equipment operation and management costs and environmental governance costs includes energy storage equipment operation and management costs and energy storage equipment environmental governance costs.
[0194] The operation and management cost of energy storage equipment is: C SV =k ch η ch P ch +k dis η dis P dis
[0195] Among them, k ch and k dis are the cost coefficients for charging and discharging operation, η ch and η dis are the charge and discharge efficiencies, P ch and P dis The power of charging and discharging respectively;
[0196] The environmental management cost of energy storage equipment is: C SVE =k SVE (P ch +P dis )
[0197] Among them, k SVEis the cost coefficient of energy storage equipment operating environment management, P ch Indicates the charging power of the energy device, P dis Expressed as the discharge power of the energy storage device;
[0198] The constraints are: P ch,min ≤P ch ≤P ch,max and P dis,min ≤P dis ≤P dis,max ;
[0199] Among them, P ch,min and P ch,max Represent the minimum and maximum power of energy storage device charging, P dis,min and P dis,max Respectively represent the minimum and maximum power of the energy storage device discharge;
[0200] The load shedding cost is: C loss =k loss P loss
[0201] Among them, k loss is the load shedding penalty factor, P loss is the load shedding power;
[0202] Constraint: 0≤P loss ≤P loss,max ;
[0203] Among them, P loss,max is the maximum allowable load shedding capacity;
[0204] The massive distributed energy optimization scheduling problem is transformed into the minimum overall power generation cost, the minimum environmental governance cost, and the minimum load shedding cost, and the objective function of the massive distributed energy optimization scheduling is obtained: min(C PV +C WT +C SV +C SVE +C loss )
[0205] The overall power balance constraint is:
[0206] Among them, P L1 is the charging constraint, P L2 is the discharge constraint condition.
[0207] Finally, the alternating direction multiplier method is used to solve the optimization problem and obtain the values of photovoltaic power generation, wind power generation, energy storage system charging or discharging power, as well as load shedding power, to achieve system optimization scheduling based on distributed energy output forecast.
[0208] The method obtains historical output data, historical environmental data, historical load data and historical scheduling strategies in the monitored area, performs orthogonal wavelet transform and reconstruction processing on the historical output data and historical environmental data respectively, and obtains a historical output reconstruction data matrix, a historical environmental reconstruction data matrix and a wind speed single-branch reconstruction data matrix, and extracts and weighted fuses the historical output reconstruction data matrix, the historical environmental reconstruction data matrix and the wind speed single-branch reconstruction data matrix through a prediction model based on global features and mutation features to obtain a first output prediction result; after solving the historical output data through an adaptive prediction model for reverse distributed energy output prediction, a second output prediction result is obtained, and the first output prediction result and the second output prediction result are compositely weightedly fused to obtain a third output prediction result. The third output prediction result is used to construct a distributed energy optimization scheduling model, so that the distributed energy optimization scheduling model solves the historical output data, historical environmental data and historical load data to obtain a scheduling result. By effectively integrating the environment, load, and scheduling, and using two-way prediction to further improve the accuracy of distributed energy output prediction, optimizing scheduling based on predicted output will make the overall scheduling of massive distributed energy more optimized.
[0209] Example 2
[0210] Accordingly, referring to FIG5 , FIG5 is a schematic diagram of the structure of the distributed energy bidirectional prediction optimization scheduling system provided by the present invention. As shown in the figure, the distributed energy bidirectional prediction optimization scheduling system includes an acquisition module 501, a first prediction module 502, a second prediction module 503, a third prediction module 504 and a scheduling result module 505, wherein the specific units of each module are as follows:
[0211] An acquisition module 501 is configured to acquire historical output data, historical environmental data, historical load data, and historical scheduling strategies within the monitored area, wherein the historical output data includes historical wind power output data and historical photovoltaic power output data;
[0212] The first prediction module 502 is configured to perform orthogonal wavelet transform and reconstruction processing on the historical output data and the historical environment data, respectively, to obtain a historical output reconstructed data matrix and a historical environment reconstructed data matrix, and perform feature extraction and weighted fusion on the historical output reconstructed data matrix and the historical environment reconstructed data matrix using a prediction model based on global features and mutation features to obtain a first output prediction result;
[0213] The second prediction module 503 is configured to perform reverse distributed energy output prediction on the historical output data, the historical load data, and the historical scheduling strategy through an adaptive prediction model to obtain a second output prediction result;
[0214] A third prediction module 504 is configured to perform a composite weighted fusion on the first output prediction result and the second output prediction result to obtain a third output prediction result;
[0215] The scheduling result module 505 is used to construct a distributed energy optimization scheduling model using the third output prediction result, and solve the distributed energy optimization scheduling model to obtain a scheduling result, so that the distribution network can schedule energy according to the scheduling result, wherein the scheduling result includes the photovoltaic power generation power value, the wind power generation power value, the energy storage system operating power value and the load shedding power value.
[0216] Optionally, the first prediction module 502 includes a normalization unit 5021, a decomposition unit 5022, a reconstruction unit 5023 and a matrix unit 5024.
[0217] The normalization unit 5021 is used to normalize each data in the historical output data and historical environmental data obtained within a preset time period to obtain normalized historical output data and normalized historical environmental data, wherein the historical environmental data includes historical ambient light intensity data, historical cut-in wind speed data, and historical cut-out wind speed data. The normalization process is as follows:
[0218] Among them, P WH Represents the historical wind power output data of renewable energy, P PVH Represents historical photovoltaic power generation data, I H Indicates the historical ambient light intensity, V ciH Indicates the historical cut-in wind speed, V coH Indicates historical cut-out wind speed, P WH,min ,P PVH,min ,P LH,min ,I H,min , V ciH,min and V coH,min and P WH,max ,P PVH,max ,I H,max , V ciH,max and V coH,max They are respectively represented by historical wind power output data, historical photovoltaic power output data, historical load data, historical ambient light intensity, historical cut-in wind speed, historical minimum and maximum cut-out wind speed, and P′ WH , P′ PVH 、I' H , V′ ciH and V′ coH They are respectively represented by the normalized historical wind power output data, historical photovoltaic power output data, historical ambient light intensity, historical cut-in wind speed, and historical cut-out wind speed;
[0219] The decomposition unit 5022 is used to perform positive wavelet transform decomposition on each data in the normalized historical output data and the normalized historical environmental data according to a preset transformation formula to obtain the real part tree decomposition wavelet coefficient and the imaginary part tree decomposition wavelet coefficient corresponding to each data in the normalized historical output data and the normalized historical environmental data. The preset transformation formula is:
[0220] Where x(t) is the signal to be processed, represents the real part tree decomposition wavelet coefficient, represents the real part tree decomposition scale coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) is the real part of the wavelet function, ψ g (t) is the imaginary wavelet function, J represents the number of layers of wavelet decomposition, and j represents the scale factor;
[0221] The reconstruction unit 5023 is used to calculate the wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal corresponding to each data according to the real part tree decomposition wavelet coefficient and the imaginary part tree decomposition wavelet coefficient through the reconstruction formula, wherein the reconstruction formula is:
[0222] Among them, d j (t) represents the reconstructed signal of a certain data wavelet coefficient, c J (t) represents the signal reconstructed by a certain data scaling coefficient, t = 1, 2, ..., M, M is the length of the signal, n is the sequence number of the signal, n = 1, 2, ..., M; represents the real part tree decomposition wavelet coefficient, represents the real part tree decomposition scale coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) represents the real part of the wavelet function, ψ g (t) represents the imaginary wavelet function, J represents the number of layers of wavelet decomposition, and j represents the scale factor;
[0223] The matrix unit 5024 is used to construct a matrix based on the wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal of each data to obtain the historical output reconstruction data matrix and the historical environment reconstruction data matrix.
[0224] Optionally, the first prediction module 502 includes a third matrix unit 5025, a convolution unit 5026 and a prediction unit 5027.
[0225] The third matrix unit 5025 is used to input the historical output reconstruction data matrix and the historical environment reconstruction data matrix into the prediction model based on global features and mutation features, and extract each data from the global trend feature layer to obtain the first matrix, extract each data from the local mutation feature layer to obtain the second matrix, and extract each data from the steady-state feature to obtain the third matrix. The calculation formula of the steady-state feature is: t =φ0+φ1x t-1 +φ2x t-2 +…+φ p x t-p +ε t
[0226] Among them, x t represents the data at time t, that is, the extracted data, φ i (i=1,…,p) is the autoregressive model coefficient, ε t Indicates that the mean is 0 and the variance is σ 2 White noise, p represents the data length of the current prediction moment value;
[0227] The convolution unit 5026 is configured to perform convolution and distillation operations on the first matrix and the second matrix respectively to obtain a new first matrix and a new second matrix;
[0228] The prediction unit 5027 is configured to fuse the new first matrix and the new second matrix to obtain a fourth matrix, and then perform weighted fusion on the third matrix and the fourth matrix to obtain a first output prediction result.
[0229] Optionally, the convolution unit 5026 includes a conversion subunit 50261, an evaluation subunit 50262, an attention value calculation subunit 50263 and a convolution subunit 50264.
[0230] The conversion subunit 50261 is used to convert the first matrix and the second matrix into vector matrices in sequence, and then multiply the first matrix by a plurality of preset weight matrices to obtain a first vector matrix, a second vector matrix, and a third vector matrix corresponding to the first matrix;
[0231] The evaluation subunit 50262 is configured to perform an approximation evaluation calculation on the first vector matrix based on a preset number of dot products to obtain an approximation evaluation standard for each vector in the first vector matrix, wherein the approximation evaluation standard is:
[0232] Among them, L Q and L K Represented as the first vector matrix and the second vector matrix The length, q iand k j are the i-th and j-th data of the first vector matrix Q1 and the second vector matrix K1, the superscript T is the transpose operation, and d represents the dimension of the matrices Q1, K1, and V1;
[0233] The attention value calculation subunit 50263 is used to obtain the contribution value of each vector to the calculation of the attention value according to the approximation evaluation standard of each vector in the first vector matrix, select the vector with the largest contribution value as the fifth matrix, and obtain the probabilistic sparse attention value according to the attention value calculation formula based on the fifth matrix, where the probabilistic sparse attention value is:
[0234] in, represents the probabilistic sparse attention value, Q1 represents the first vector matrix, K1 represents the second vector matrix, V1 represents the third vector matrix, d is the dimension of the matrices Q1, K1, and V1, and the superscript T is the transpose operation;
[0235] The convolution subunit 50264 is used to perform convolution and distillation operations according to the probabilistic sparse attention value and the preset weight to obtain a new first matrix and a new second matrix, wherein the convolution and distillation operation process is:
[0236] in,[·] AB Represents the value obtained after an attention module, Conv1d is a one-dimensional convolution, ELU is the activation function of the activation layer, MaxPool is the maximum pooling, is time t The data after the j+1th distillation layer is the new first matrix.
[0237] Optionally, the second prediction module 503 includes an expansion unit 5031 and a second prediction unit 5032.
[0238] The expansion unit 5031 is used to expand the historical scheduling strategy to obtain a new historical scheduling strategy;
[0239] The second prediction unit 5032 is used to construct an adaptive prediction model based on historical load data, a new historical scheduling strategy, historical cut-in wind speed, historical cut-out wind speed and historical ambient light intensity, and solve the adaptive prediction model to obtain a second output prediction result, wherein the adaptive prediction model includes a wind adaptive prediction model and a photovoltaic adaptive prediction module, and the second output prediction result includes a wind output prediction result and a photovoltaic output prediction result.
[0240] Optionally, the second prediction module 503 is further configured to construct an adaptive prediction model based on historical load data, a new historical scheduling strategy, historical cut-in wind speed, historical cut-out wind speed, and historical ambient light intensity, and solve the adaptive prediction model to obtain a second output prediction result, specifically:
[0241] A wind adaptive prediction model is constructed based on historical load data, new historical dispatching strategies, historical cut-in wind speeds, and historical cut-out wind speeds. The wind adaptive prediction model is: P WH =β Wi X Wi +β W
[0242] Among them, P WH represents the wind power output prediction result, β Wi (i=1,2,3,4) is the regression coefficient between wind power output and related factors, β W is the overall error of the wind output model, X Wi represents the wind power output prediction factor matrix, X Wi =[P LH ,D H ,V ciH ,V coH ], where P LH Indicates historical load data, D H Represents the new historical scheduling strategy, V ciH Indicates the historical cut-in wind speed and V coH Historical cut-out wind speed;
[0243] Solve P using the least squares method WH =β Wi X Wi +β W Regression coefficient, get the initial value of regression coefficient Further introduction of weight Then the weight vector It can be expressed as:
[0244] renew x j It's X Wi Variables in
[0245] For all regularization factors λ n , solve
[0246] calculate Will Substitute P WH =β Wi X Wi +β W, that is, the wind power output prediction result is obtained;
[0247] For photovoltaic output, an adaptive prediction model is established based on historical load, new historical scheduling strategy, and historical ambient light intensity: PVH =β PVi X PVi +β PV
[0248] Among them, P PVH represents the photovoltaic output regression model, β PVi (i=1,2,3) is the regression coefficient between photovoltaic output and related factors, β PV is the overall error of the photovoltaic output model, X PVi represents the photovoltaic output prediction factor matrix, X PVi =[P LH ,D H ,I H ], where P LH Indicates historical load data, D H Represents the new historical scheduling strategy and I H Indicates the historical ambient light intensity;
[0249] Solve P using the least squares method PVH =β PVi X PVi +β PV The regression coefficient is obtained by Further introduction of weight Then the weight vector It can be expressed as:
[0250] renew x j It's X PVi Variables in
[0251] For all regularization factors λ n , solve
[0252] calculate Will Substitute P PVH =β PVi X PVi +β PV , get the photovoltaic output prediction result;
[0253] Optionally, the third prediction module 504 includes a prediction unit 5041,
[0254] The prediction unit 5041 is configured to perform a composite weighted fusion on the first output prediction result and the second output prediction result to obtain a third output prediction result, specifically:
[0255] The first output prediction result and the second output prediction result are combined by weighted average method to obtain a third output prediction result. The specific calculation formula is: HP =θP HP1 +(1-θ)P HP2
[0256] Among them, P HP Indicates the third output prediction result, P HP1 Indicates the first output prediction result, P HP2 represents the second output prediction result, and θ represents the power coefficient predicted by the forward prediction method.
[0257] The scheduling result module 505 includes a power generation cost unit 5051, an energy storage device operation and management cost 5052, a load shedding cost 5053, a scheduling model construction unit 5054 and a calculation unit 5055.
[0258] The power generation cost unit 5051 is used to construct according to the third prediction result and the power generation cost coefficient to obtain the power generation cost, wherein the power generation cost is the photovoltaic power generation cost or the wind power generation cost;
[0259] Energy storage equipment operation and management cost 5052 is used to construct the energy storage equipment operation and management cost based on the charging operation cost coefficient, the discharging operation cost coefficient, the charging efficiency, the discharging efficiency, the charging power, and the discharging power;
[0260] The load shedding cost 5053 is constructed based on the energy storage equipment operating environment management cost coefficient, the energy storage equipment charging power, and the energy storage equipment discharging power to obtain the energy storage equipment environmental management cost, and is constructed based on the load shedding penalty coefficient and the load shedding power to obtain the load shedding cost;
[0261] The scheduling model construction unit 5054 is used to construct the objective function and constraint conditions of the distributed energy optimization scheduling model based on the power generation cost, the energy storage equipment operation and management cost, the energy storage equipment environmental management cost and the load shedding cost, wherein the objective function is: min(C PV +C WT +C SV +C SVE +C loss )
[0262] Among them, C PV is the photovoltaic power generation cost, C WT is the cost of wind power generation, C SV is the operation and management cost of energy storage equipment, CSVE is the environmental management cost of energy storage equipment, C loss is the load shedding cost;
[0263] The calculation unit 5055 is used to solve the objective function and the constraint conditions using the alternating direction multiplier method to obtain the scheduling result, wherein the scheduling result includes the photovoltaic power generation power value, the wind power generation power value, the energy storage system operating power value and the load shedding power value.
[0264] The more detailed working principle and process flow of this embodiment can be referred to, but not limited to, the relevant records of the first embodiment.
[0265] Compared with the existing technology, by obtaining historical output data, historical environmental data, historical load data and historical scheduling strategies in the monitored area, orthogonal wavelet transform and reconstruction processing are performed on the historical output data and historical environmental data respectively to obtain the historical output reconstruction data matrix, the historical environmental reconstruction data matrix and the wind speed single-branch reconstruction data matrix, and the historical output reconstruction data matrix, the historical environmental reconstruction data matrix and the wind speed single-branch reconstruction data matrix are subjected to feature extraction and weighted fusion through a prediction model based on global features and mutation features to obtain a first output prediction result; after the historical output data is subjected to reverse distributed energy output prediction solution through an adaptive prediction model, a second output prediction result is obtained, and a composite weighted fusion is performed on the first output prediction result and the second output prediction result to obtain a third output prediction result, and a distributed energy optimization scheduling model is constructed using the third output prediction result, so that the distributed energy optimization scheduling model solves the historical output data, historical environmental data and historical load data to obtain a scheduling result. This method effectively integrates the environment, load, and scheduling, uses two-way prediction to further improve the accuracy of distributed energy output prediction, and establishes optimized scheduling based on predicted output, making the overall scheduling of massive distributed energy more optimized.
[0266] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.
Claims
1. A distributed energy bidirectional prediction optimization scheduling method, characterized in that: include: S1: Acquire historical output data, historical environmental data, historical load data, and historical dispatch strategies in the area to be monitored, wherein the historical output data includes historical wind power output data and historical photovoltaic power output data; S2: performing orthogonal wavelet transform and reconstruction processing on the historical output data and the historical environment data respectively to obtain a historical output reconstruction data matrix and a historical environment reconstruction data matrix, and extracting features and weighted fusion of the historical output reconstruction data matrix and the historical environment reconstruction data matrix through a prediction model based on global features and local mutations to obtain a first output prediction result; S3: After performing reverse distributed energy output prediction and solving the historical output data, the historical load data and the historical dispatching strategy through an adaptive prediction model, a second output prediction result is obtained; S4: performing composite weighted fusion on the first output prediction result and the second output prediction result to obtain a third output prediction result; S5: Utilize the third output prediction result to construct a distributed energy optimization scheduling model, and solve the distributed energy optimization scheduling model to obtain a scheduling result, so that the distribution network can schedule energy according to the scheduling result, wherein the scheduling result includes photovoltaic power generation value, wind power generation value, energy storage system operating power value and load shedding power value.
2. A distributed energy bidirectional prediction optimization scheduling method as claimed in claim 1, characterized in that: The historical output data and the historical environment data are respectively subjected to orthogonal wavelet transformation and reconstruction processing to obtain a historical output reconstruction data matrix and a historical environment reconstruction data matrix, which are specifically: Each of the historical output data and the historical environment data acquired within a preset time period is normalized to obtain normalized historical output data and normalized historical environment data, wherein the historical environment data includes historical environment light intensity data, historical cut-in wind speed data, and historical cut-out wind speed data, wherein the normalization process is: Among them, P WH Represents the historical wind power output data of renewable energy, P PVH Represents historical photovoltaic power generation output data, I H Represents the historical ambient light intensity, V ciH represents the historical cut-in wind speed, V coH Indicates the historical cut-out wind speed, P WH,min ,P PVH,min ,I H,min , V ciH,min and V coH,min and P WH,max ,P PVH,max ,I H,max , V ciH,max and V coH,max They are respectively represented by historical wind power output data, historical photovoltaic power output data, historical ambient light intensity, historical cut-in wind speed, historical minimum and maximum cut-out wind speed, and P′ WH , P′ PVH , I′ H , V′ ciH and V′ coH They are respectively represented by the normalized historical wind power output data, historical photovoltaic power output data, historical ambient light intensity, historical cut-in wind speed, and historical cut-out wind speed; The normalized historical output data and the normalized historical environmental data are respectively decomposed by positive wavelet transform according to a preset transformation formula to obtain the real part tree decomposition wavelet corresponding to each data in the normalized historical output data and the normalized historical environmental data. The wavelet coefficients and the imaginary part tree decomposition wavelet coefficients, wherein the preset transformation formula is: Among them, x(t) is the signal to be processed, represents the real part of the tree decomposition wavelet coefficient, represents the real part tree decomposition scaling coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) is the real part of the wavelet function, ψ g (t) is the imaginary wavelet function, J represents the number of layers of wavelet decomposition, and j represents the scale factor; The wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal corresponding to each data are obtained by calculating the wavelet coefficients decomposed by the real tree and the wavelet coefficients decomposed by the imaginary tree through a reconstruction formula, wherein the reconstruction formula is: Among them, d j (t) represents the reconstructed signal of a certain data wavelet coefficient, c J (t) represents the reconstructed signal of a certain data scale coefficient, t = 1, 2, ..., M, M is the length of the reconstructed signal, n is the sequence number of the reconstructed signal, n = 1, 2, ..., M; represents the real part of the tree decomposition wavelet coefficient, represents the real part tree decomposition scaling coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) represents the real part of the wavelet function, ψ g (t) represents the imaginary wavelet function, J represents the number of layers of wavelet decomposition, and j represents the scale factor; A matrix is constructed according to the wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal of each data to obtain a historical output reconstruction data matrix and a historical environment reconstruction data matrix.
3. A distributed energy bidirectional prediction optimization scheduling method as claimed in claim 1, characterized in that: The historical output reconstruction data matrix and the historical environment reconstruction data matrix are subjected to feature extraction and weighted fusion through a prediction model based on global trends and local mutations to obtain a first output prediction result, specifically: The historical output reconstruction data matrix and the historical environment reconstruction data matrix are input into the prediction model based on global features and mutation features, and the global trend feature layer is used to extract each data to obtain the first matrix, and the local mutation feature layer is used to extract each data to obtain the second matrix. The steady-state feature is used to extract each data to obtain the third matrix, wherein the steady-state feature calculation formula is: x t =φ0+φ1x t-1 +φ2x t-2 +…+φ p x t-p +e t Among them, x t represents the data at time t, that is, the extracted data, φ i (i=1,…,p) is the autoregressive model coefficient, ε t Indicates that the mean is 0 and the variance is σ 2 white noise, p represents the data length of the current prediction time value; after performing convolution and distillation operations on the first matrix and the second matrix respectively, a new first matrix and a new second matrix are obtained; After the new first matrix and the new second matrix are fused to obtain a fourth matrix, the third matrix and the fourth matrix are weightedly fused to obtain a first output prediction result.
4. A distributed energy bidirectional prediction optimization scheduling method as claimed in claim 3, characterized in that: After performing convolution and distillation operations on the first matrix and the second matrix respectively, a new first matrix and a new second matrix are obtained, which are specifically: After converting the first matrix and the second matrix into vector matrices in sequence, multiplying the first matrix by a plurality of preset weight matrices respectively to obtain a first vector matrix, a second vector matrix, and a third vector matrix corresponding to the first matrix; An approximation evaluation calculation is performed on the first vector matrix according to a preset number of dot products to obtain an approximation evaluation standard for each vector in the first vector matrix, wherein the approximation evaluation standard is: Among them, L Q and L K Respectively represented as the first vector matrix and the second vector matrix The length, q i and k j are the i-th and j-th data of the first vector matrix Q1 and the second vector matrix K1, the superscript T is the transposition operation, and d represents the dimension of the matrices Q1, K1, and V1; According to the approximation evaluation standard of each vector in the first vector matrix, the contribution value of each vector to the calculation of the attention value is obtained, the vector with the largest contribution value is selected as the fifth matrix, and the probabilistic sparse attention value is obtained according to the fifth matrix through the attention value calculation formula, wherein the probabilistic sparse attention value is: in, represents the probabilistic sparse attention value, Q1 represents the first vector matrix, K1 represents the second vector matrix, V1 represents the third vector matrix, d is the dimension of the matrices Q1, K1, and V1, and the superscript T is the transpose operation; After performing convolution and distillation operations according to the probabilistic sparse attention value and the preset weight, a new first matrix and a new second matrix are obtained, wherein the convolution and distillation operation process is: in,[·] AB Represents the value obtained after an attention module, Conv1d is a one-dimensional convolution, ELU is the activation function of the activation layer, and MaxPool is the maximum pooling. is time t The data after the j+1th distillation layer is the new first matrix.
5. A distributed energy bidirectional prediction optimization scheduling method as claimed in claim 1, characterized in that: After the historical output data, the historical load data and the historical dispatching strategy are used to perform reverse distributed energy output prediction and solution through the adaptive prediction model, a second output prediction result is obtained, which is specifically: Expanding the historical scheduling strategy to obtain a new historical scheduling strategy; An adaptive prediction model is constructed based on the historical load data, the new historical scheduling strategy, the historical cut-in wind speed, the historical cut-out wind speed and the historical ambient light intensity, and the adaptive prediction model is solved to obtain a second output prediction result, wherein the adaptive prediction model includes a wind adaptive prediction model and a photovoltaic adaptive prediction module, and the second output prediction result includes a wind output prediction result and a photovoltaic output prediction result.
6. A distributed energy bidirectional prediction optimization scheduling method as claimed in claim 5, characterized in that: The adaptive prediction model is constructed according to the historical load data, the new historical dispatching strategy, the historical cut-in wind speed, the historical cut-out wind speed and the historical ambient light intensity, and the adaptive prediction model is solved to obtain the second output prediction result, which is specifically: A wind power adaptive dispatch strategy is constructed based on the historical load data, the new historical dispatch strategy, the historical cut-in wind speed and the historical cut-out wind speed. The wind power adaptive prediction model is: P WH =b Wi X Wi +b W Among them, P WH represents the wind power output forecast result, β Wi (i=1,2,3,4) is the regression coefficient between wind power output and related factors, β W is the overall error of the wind output model, X Wi represents the wind power output prediction factor matrix, X Wi =[P LH ,D H ,V ciH ,V coH ], where P LH Indicates historical load data, D H represents the new historical scheduling strategy, V ciH Indicates the historical cut-in wind speed and V coH Historical cut-out wind speed; Solve P using the least squares method WH =β Wi X Wi +β W Regression coefficient, get the initial value of regression coefficient Further introduction of weight The weight vector It can be expressed as: renew x j Yes X Wi The variables in For all regularization factors λ n , solve calculate Will Substitute P WH =β Wi X Wi +β W , get the wind power output prediction result; For photovoltaic output, an adaptive prediction model based on historical load, new historical scheduling strategy, and historical ambient light intensity is established: P PVH =b PVi X PVi +b PV Among them, P PVH represents the photovoltaic output regression model, β PVi (i=1,2,3) is the regression coefficient between photovoltaic output and related factors, β PV is the overall error of the photovoltaic output model, X PVi represents the photovoltaic output prediction factor matrix, X PVi =[P LH ,D H ,I H ], where P LH Indicates historical load data, D H Represents the new historical scheduling strategy and I H Indicates the historical ambient light intensity; Solve P using the least squares method PVH =β PVi X PVi +β PV The regression coefficient is obtained by Further introduction of weight The weight vector It can be expressed as: renew x j Yes X PVi The variables in For all regularization factors λ n , solve calculate Will Substitute P PVH =β PVi X PVi +β PV , that is, the photovoltaic output prediction result is obtained.
7. A distributed energy bidirectional prediction optimization scheduling method according to claim 1, characterized in that: The first output prediction result and the second output prediction result are compositely weighted and fused to obtain a third output prediction result, specifically: The first output prediction result and the second output prediction result are combined by weighted average method to obtain a third output prediction result. The specific calculation formula is: P HP =θP HP1 +(1-θ)P HP2 Among them, P HP represents the third output prediction result, P HP1 Represents the first output prediction result, P HP2 represents the second output prediction result, and θ represents the power coefficient predicted by the forward prediction method.
8. A distributed energy bidirectional prediction optimization scheduling method as claimed in claim 1, characterized in that: The distributed energy optimization scheduling model is constructed by using the third output prediction result, and the distributed energy optimization scheduling model is used to optimize the historical output data and the energy storage device operation data to obtain the scheduling result, specifically: Constructing according to the third prediction result and the power generation cost coefficient to obtain the power generation cost, wherein the power generation cost is the photovoltaic power generation cost or the wind power generation cost; The operation and management cost of the energy storage equipment is obtained by constructing the charging operation cost coefficient, the discharging operation cost coefficient, the charging efficiency, the discharging efficiency, the charging power and the discharging power; The environmental governance cost of the energy storage equipment is obtained by constructing the cost coefficient of the energy storage equipment operating environment governance, the energy storage equipment charging power and the energy storage equipment discharging power, and the load shedding cost is obtained by constructing the load shedding penalty coefficient and the load shedding power; The objective function and constraint conditions of the distributed energy optimization scheduling model are constructed according to the power generation cost, the energy storage equipment operation and management cost, the energy storage equipment environmental management cost and the load shedding cost, wherein the objective function is: min(C PV +C WT +C SV +C SVE +C loss ) Among them, C PV is the photovoltaic power generation cost, C WT is the cost of wind power generation, C SV is the operation and management cost of energy storage equipment, C SVE is the environmental management cost of energy storage equipment, C loss is the load shedding cost; The objective function and constraint conditions are solved using the alternating direction multiplier method to obtain a scheduling result, wherein the scheduling result includes a photovoltaic power generation value, a wind power generation value, an energy storage system operating power value, and a load shedding power value.
9. A distributed energy bidirectional prediction optimization scheduling system, characterized in that: include: An acquisition module, used to acquire historical output data, historical environmental data, historical load data and historical dispatching strategies in the area to be monitored, wherein the historical output data includes historical wind power output data and historical photovoltaic power output data; A first prediction module is used to perform orthogonal wavelet transformation and reconstruction processing on the historical output data and the historical environment data, respectively, to obtain a historical output reconstruction data matrix and a historical environment reconstruction data matrix, and to perform feature extraction and weighted fusion on the historical output reconstruction data matrix and the historical environment reconstruction data matrix through a prediction model based on global trends and local mutations, to obtain a first output prediction result; A second prediction module is used to perform reverse distributed energy output prediction and solution on the historical output data, the historical load data and the historical dispatching strategy through an adaptive prediction model to obtain a second output prediction result; A third prediction module, configured to perform a composite weighted fusion on the first output prediction result and the second output prediction result to obtain a third output prediction result; A scheduling result module is used to use the third output prediction result to build a distributed energy optimization scheduling model, and use the distributed energy optimization scheduling model to optimize the historical output data and energy storage equipment operation data to obtain a scheduling result, so that the distribution network can schedule energy according to the scheduling result, wherein the scheduling result includes photovoltaic power generation power value, wind power generation power value, energy storage system operation power value and load shedding power value.
10. The distributed energy bidirectional prediction optimization scheduling system according to claim 9, characterized in that: The first prediction module includes a normalization unit, a decomposition unit, a reconstruction unit and a matrix unit. The normalization unit is used to normalize each data in the historical output data and the historical environment data acquired within a preset time period to obtain normalized historical output data and normalized historical environment data, wherein the historical environment data includes historical environment light intensity data, historical cut-in wind speed data, and historical cut-out wind speed data, wherein the normalization process is: Among them, P WH Represents the historical wind power output data of renewable energy, P PVH Represents historical photovoltaic power generation output data, I H Represents the historical ambient light intensity, V ciH represents the historical cut-in wind speed, V coH Indicates the historical cut-out wind speed, P WH,min ,P PVH,min ,I H,min , V ciH,min and V coH,min and P WH,max ,P PVH,max ,P LH,max ,I H,max , V ciH,max and V coH,max They are respectively represented by historical wind power output data, historical photovoltaic power output data, historical ambient light intensity, historical cut-in wind speed, historical minimum and maximum cut-out wind speed, and P′ WH , P′ PVH , I′ H , V′ ciH and V′ coH They are respectively represented by normalized historical wind power output data, historical photovoltaic power output data, historical ambient light intensity, historical cut-in wind speed, and historical cut-out wind speed; The decomposition unit is used to perform positive wavelet transform decomposition on each data in the normalized historical output data and the normalized historical environment data according to a preset transformation formula, to obtain the real tree decomposition wavelet coefficient and the imaginary tree decomposition wavelet coefficient corresponding to each data in the normalized historical output data and the normalized historical environment data, wherein the preset transformation formula is: Among them, x(t) is the signal to be processed, represents the real part of the tree decomposition wavelet coefficient, represents the real part tree decomposition scaling coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) is the real part of the wavelet function, ψ g (t) is the imaginary wavelet function, J represents the number of layers of wavelet decomposition, and j represents the scale factor; The reconstruction unit is used to calculate the wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal corresponding to each data through a reconstruction formula according to the real part tree decomposition wavelet coefficient and the imaginary part tree decomposition wavelet coefficient, wherein the reconstruction formula is: Among them, d j (t) represents the reconstructed signal of a certain data wavelet coefficient, c J (t) represents the signal reconstructed by a certain data scale coefficient, t = 1, 2, ..., M, M is the length of the signal, n is the sequence number of the signal, n = 1, 2, ..., M; represents the real part of the tree decomposition wavelet coefficient, represents the real part tree decomposition scaling coefficient, represents the imaginary tree decomposition wavelet coefficients, represents the imaginary tree decomposition scaling coefficient, ψ h (t) represents the real part of the wavelet function, ψ g (t) represents the imaginary wavelet function, J represents the number of layers of wavelet decomposition, and j represents the scale factor; The matrix unit is used to construct a matrix according to the wavelet coefficient reconstruction signal and the scale coefficient reconstruction signal of each data, so as to obtain a historical output reconstruction data matrix and a historical environment reconstruction data matrix.
Citation Information
Patent Citations
Regional power grid dispatching system and method based on prediction control
CN105337415A
Economic dispatching method for wind power and photovoltaic combined power generation system
CN110880789A
Multi-region energy demand prediction method and device, terminal equipment and storage medium
CN113627677A
Energy management system for wind-light-fire-hydrogen storage new energy system
CN113922375A
Generated power prediction device, method for predicting generated power, and program for predicting generated power
JP2014217092A
Cited By
Distributed new energy grid-connected data processing method and system
CN119253743A
Intelligent monitoring method and system for full-process progress of agricultural machinery part production
CN119294789A
User side energy storage operation monitoring analysis method based on source network load cooperation
CN119382134A
Distributed photovoltaic output power estimation method
CN119401450A
Wind power prediction method and device, electronic equipment and storage medium
CN119419801A
