Intelligent power grid power dispatching optimization method

Abstract

The invention relates to the technical field of smart power grids, and discloses a smart power grid power dispatching optimization method, which comprises the following steps of: firstly, acquiring power grid operation data, and processing data missing and noise problems by utilizing federal learning; and constructing a load prediction model through a dynamic time warping algorithm and a specific network. A multi-energy coupling scheduling model and a demand response game model are constructed, and a multi-time scale rolling optimization framework is established. And carrying out sensitivity analysis on scheduling parameters, designing a hierarchical collaborative optimization mechanism, and constructing a robust optimization model to cope with the power flow uncertainty. And integrating a scheduling instruction verification module, and deploying an online incremental learning mechanism. The method can effectively process data, accurately predict load, optimize multi-energy scheduling, guide demand response, deal with uncertainty, verify scheduling instructions and update the model in real time, improves the safety, reliability and economy of smart grid power scheduling, and realizes optimal configuration of power resources.

Description

Technical Field

[0001] The present invention relates to the technical field of smart grids, and specifically to a method for optimizing power dispatching in a smart grid. Background Art

[0002] With the energy transition and the growth of power demand, the construction of smart grids has become a key direction for the development of the power industry. Traditional power grid power dispatching methods face many challenges when dealing with the characteristics of new power systems. In terms of data processing, the data sources of smart grids are extensive and complex, including data from power generation units, distribution networks, energy storage devices, renewable energy sources, and user loads. These data are prone to missing or noise problems. The traditional centralized data processing method is not only inefficient but also has a risk of privacy leakage. For example, when the power data of multiple regions are centrally processed, it is difficult to guarantee the data privacy of each region, and large-scale data transmission will cause communication congestion, affecting the timeliness of data processing and unable to meet the real-time requirements of smart grids.

[0003] Power load forecasting is an important basis for power dispatching. Traditional forecasting methods are difficult to accurately capture the load fluctuation characteristics and temporal correlation, resulting in low forecasting accuracy. In smart grids, meteorological factors have a significant impact on the load, while traditional methods do not fully consider the spatio-temporal coupling relationship between meteorological factors and the load. For example, in extreme weather conditions, traditional forecasting models cannot quickly and accurately reflect the load changes, leading to a mismatch between power dispatching and actual demand, increasing the uncertainty and cost of power supply.

[0004] Multi-energy coupled dispatching is an important feature of smart grids. When multiple energy sources such as thermal power, wind power, photovoltaic power, and energy storage are coordinated for dispatching, the power flow constraints between energy nodes are complex, and traditional dispatching models cannot effectively handle them. It is also difficult to accurately calculate the energy transmission losses across voltage levels, resulting in low energy utilization efficiency. For example, in areas rich in wind and solar resources, due to the lack of an effective dispatching model, the phenomenon of wind and light abandonment occurs frequently, causing energy waste.

[0005] In terms of demand response, the user load adjustment behavior is complex under real-time electricity price fluctuations. Traditional electricity price strategies do not fully consider the game characteristics of user behavior and are difficult to effectively guide users to reasonably adjust their electricity consumption behavior. This makes it difficult to achieve power supply-demand balance through flexible regulation on the demand side, reducing the overall operating efficiency of the power system.

[0006] In addition, the access of distributed power sources increases the uncertainty of power grid power flow, and traditional deterministic optimization methods cannot cope with it. Moreover, there is a lack of effective verification means for the security and reliability of dispatching instructions, and there are potential risks in the operation of the power grid. The traditional power dispatching optimization model is not updated in a timely manner and cannot adapt to the rapidly changing operating state of the power grid. Summary of the Invention

[0007] The purpose of the present invention is to provide an intelligent power grid power dispatching optimization method to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solution: an intelligent power grid power dispatching optimization method, the method comprising:

[0009] Step 1: Obtain power grid operation data, including generator output curves, distribution network topological structures, energy storage device capacities, renewable energy predicted outputs, and user load time series data. When the data is missing or noisy, use the federated learning method to jointly perform distributed data cleaning and imputation with multi-region edge nodes;

[0010] Step 2: Align historical load data and meteorological factor time series based on the dynamic time warping algorithm, construct a power load spatio-temporal coupling prediction model, use a deep residual network enhanced by an attention mechanism to extract load fluctuation features, fuse the LSTM network to capture temporal correlations, and generate regional-level load prediction results;

[0011] Step 3: Construct a multi-energy coupling dispatching model, with thermal power, wind power, photovoltaic power, and energy storage as decision-making units, use a double-layer graph convolutional network to model the power flow constraints between energy nodes, define the node balance equation in the form of the tensor product of the topological adjacency matrix and the energy conversion efficiency, and introduce virtual nodes to handle the energy transmission losses across voltage levels;

[0012] Step 4: Establish a demand response game model based on real-time electricity price fluctuation signals, model the user load adjustment behavior as a non-cooperative game process under incomplete information, use the quantum particle swarm algorithm to solve the Nash equilibrium point, and generate a time-of-use electricity price incentive strategy;

[0013] Step 5: Establish a multi-time scale rolling optimization framework, decompose the dispatching period into three stages: day-ahead, intra-day, and real-time, use the deep reinforcement learning algorithm to construct a dynamic decision-making strategy, define the state space as the power grid operation parameters, the action space as the unit start-stop instructions and the energy storage charge-discharge rate, and update the policy network parameters through the double delayed deep deterministic policy gradient algorithm;

[0014] Step 6: Conduct a multi-objective sensitivity analysis of the dispatching parameters, use the improved Hilbert-Schmidt independence criterion to screen key variables, construct a surrogate model based on sparse polynomial chaos expansion to replace the high-dimensional optimization problem, and use adaptive sparse grid sampling to improve the model accuracy;

[0015] Step 7: Design a hierarchical collaborative optimization mechanism, use mixed integer programming in the outer layer to optimize the unit combination and network topology, use the alternating direction multiplier method in the inner layer to solve the power distribution problem, and achieve the iterative convergence of cross-layer coupling constraints by transmitting boundary conditions through Lagrange multipliers;

[0016] Step 8: In view of the power flow uncertainty caused by the access of distributed power sources, a robust optimization model is constructed, interval fuzzy sets are used to describe the fluctuation range of wind and solar power output, the robustness index is defined as the system stability threshold under the worst scenario, and the main problem and sub-problems are decoupled and solved based on the Benders decomposition algorithm;

[0017] Step 9: Integrate the scheduling instruction verification module, use smart contract technology to perform topological security verification and N-1 fault rehearsal on the optimization results, generate scheduling instruction execution certificates through formal verification methods, and implement cross-domain instruction synchronization under privacy protection based on zero-knowledge proof;

[0018] Step 10: Deploy an online incremental learning mechanism, use an online sequence extreme learning machine to update the load forecasting model parameters in real time, combine the sliding window technology to dynamically adjust the optimization model weight coefficients, and evaluate the model drift based on the Markov chain Monte Carlo method to trigger the retraining conditions.

[0019] Preferably, the federated learning adopts a dynamic weight aggregation strategy to calculate the aggregation weight according to the edge node data quality assessment value. The data quality assessment indicators include data integrity, time series continuity and KL divergence with the central node distribution, and a gradient transmission protocol based on Paillier homomorphic encryption is designed to prevent leakage of original data.

[0020] Preferably, the deep residual network comprises a residual block stacking structure, each residual block is composed of a gated convolution layer, a batch normalization layer and a jump connection, the activation function of the gated convolution layer adopts an adaptive gated activation function, and the network output end is connected to the self-attention module to calculate the spatial correlation weight matrix of the load characteristics.

[0021] Preferably, the particle position encoding in the quantum particle swarm algorithm is represented by quantum bit phase angle, the quantum rotating gate update strategy is defined as the phase difference function between the particle historical optimum and the global optimum, and quantum entanglement operation is introduced to realize particle swarm collaborative search, and the algorithm termination condition is that the group diversity is lower than a preset threshold or reaches the maximum number of iterations.

[0022] Preferably, the basis function of the sparse polynomial chaos expansion proxy model selects Legendre orthogonal polynomials, an adaptive forward-backward selection algorithm is used to determine the polynomial order and cross terms, and the model complexity is optimized based on the Akaike information criterion.

[0023] Preferably, the main problem of the Benders decomposition algorithm is modeled as a mixed integer quadratic programming problem, and the subproblems are converted into semidefinite programming forms to solve the extreme value of system stability, and an accelerated cutting plane generation strategy is designed, and dual variable sensitivity analysis is used to screen effective constraints.

[0024] Preferably, the formal verification method uses temporal logic formulas to describe power grid security constraints, including bus voltage over-limit, line overload, and frequency deviation indicators, and verifies the compliance of dispatching instructions by traversing the state transition paths through symbolic model checking technology.

[0025] Preferably, the hidden layer node dynamic expansion strategy of the online sequential extreme learning machine adopts an error sensitivity criterion, defines the node addition threshold as the moving average variance of the prediction residuals, and designs a forgetting factor mechanism to attenuate the influence weight of historical data on model update.

[0026] Preferably, the intelligent contract verification module includes a topology connectivity verification sub-module, which uses a depth-first search algorithm to detect the risk of island operation, and calculates the minimum break point set based on an improved Floyd-Warshall shortest path algorithm to ensure the satisfaction rate of the N-1 criterion.

[0027] Preferably, the Markov chain Monte Carlo method uses a sampler to generate the posterior distribution of model parameters, defines the drift amount index as the distance of the parameter distribution, and triggers the full model retraining process when the drift amount exceeds a preset threshold.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] Distributed data cleaning and imputation are carried out by using federated learning to jointly process multi-region edge nodes. The dynamic weight aggregation strategy is used to improve the accuracy of data processing. The gradient transmission protocol combined with Paillier homomorphic encryption is used to prevent the leakage of original data, which not only protects data privacy but also can effectively process large-scale complex data, improve data quality, and provide a reliable basis for subsequent dispatching decisions. For example, power data in different regions can be cleaned and imputed locally, and only encrypted gradients are transmitted, avoiding the leakage of sensitive data and improving the accuracy of data fusion at the same time.

[0030] Based on the dynamic time warping algorithm to align historical load data with meteorological factor time series, the constructed power load spatio-temporal coupling prediction model uses a deep residual network and an LSTM network enhanced by an attention mechanism, which can more accurately extract load fluctuation characteristics and capture temporal correlations, significantly improving the accuracy of regional-level load prediction. In practical applications, it can effectively reduce the problem of power supply-demand imbalance caused by load prediction errors and reduce the operating cost of the power system.

[0031] The multi-energy coupling scheduling model takes thermal power, wind power, photovoltaic power, and energy storage as decision-making units, uses a double-layer graph convolutional network to model the power flow constraints between energy nodes, introduces virtual nodes to handle the energy transmission losses across voltage levels, realizes the efficient collaborative scheduling of multiple energies, and improves the energy utilization efficiency. For example, by accurately calculating power flow and losses, energy waste can be reduced, the ability of the power grid to accommodate renewable energy can be improved, and the green transformation of the energy structure can be promoted.

[0032] The demand response game model established based on real-time electricity price fluctuation signals models the user's load adjustment behavior as a non-cooperative game process under incomplete information. The quantum particle swarm algorithm is used to solve the Nash equilibrium point to generate a time-of-use electricity price incentive strategy, which can effectively guide users to reasonably adjust their electricity consumption behavior, achieve a dynamic balance between electricity supply and demand, and improve the flexibility and stability of the power system.

[0033] The multi-time-scale rolling optimization framework decomposes the dispatch cycle into three stages: day-ahead, intraday, and real-time. A deep reinforcement learning algorithm is used to build a dynamic decision-making strategy. The policy network parameters are updated through a double-delayed deep deterministic policy gradient algorithm. The dispatch strategy can be quickly adjusted according to the real-time operation status of the power grid, improving the real-time and adaptability of the dispatch decision, and ensuring the stable operation of the power grid at different time scales. Multi-objective sensitivity analysis is performed on the dispatch parameters, and the improved Hilbert-Schmidt independence criterion is used to screen key variables. An agent model based on sparse polynomial chaos expansion is constructed to replace high-dimensional optimization problems. The model accuracy is improved through adaptive sparse grid sampling, which can greatly reduce the optimization calculation complexity and improve the optimization efficiency, providing an efficient calculation method for the optimization dispatch of complex power systems. The designed hierarchical collaborative optimization mechanism uses mixed integer programming to optimize the unit combination and network topology in the outer layer, and the alternating direction multiplier method is used in the inner layer to solve the power allocation problem. The iterative convergence of cross-layer coupling constraints is achieved through the Lagrange multiplier transfer boundary conditions. It can effectively coordinate the optimization decisions at different levels of the power system, improve the overall optimization effect, and reduce the system operation cost.

[0034] In view of the power flow uncertainty caused by the access of distributed power sources, the robust optimization model constructed uses interval fuzzy sets to describe the fluctuation range of wind and solar power output, defines the robustness index as the system stability threshold under the worst scenario, and decouples and solves it based on the Benders decomposition algorithm, which enhances the grid's ability to cope with uncertainty and improves the safety and reliability of grid operation. The integrated dispatch instruction verification module uses smart contract technology to perform topological security verification and N-1 fault rehearsal, generates dispatch instruction execution certificates through formal verification methods, and implements cross-domain instruction synchronization under privacy protection based on zero-knowledge proof, ensuring the security and reliability of dispatch instructions while protecting data privacy in cross-domain dispatching.

[0035] An online incremental learning mechanism is deployed, and an online sequence extreme learning machine is used to update the load forecasting model parameters in real time. The sliding window technology is combined to dynamically adjust the optimization model weight coefficients, and the Markov chain Monte Carlo method is used to evaluate the model drift to trigger the retraining conditions. This enables the model to track changes in the operating status of the power grid in real time, continuously maintain good prediction and optimization performance, and improve the accuracy and adaptability of power dispatching. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1This is the working principle diagram of the intelligent power grid power dispatching optimization method of the present invention;

[0037] Figure 2 This is a schematic diagram of federated learning data processing and privacy protection;

[0038] Figure 3 This is a schematic diagram of using the Benders decomposition algorithm to solve the robust optimization model. Specific implementation manners

[0039] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0040] Please refer to Figures 1-3 , the present invention provides an intelligent power grid power dispatching optimization method, and its overall implementation scheme is as follows:

[0041] Step 1: Obtain various key data of power grid operation, such as generator output curves, distribution network topology structures, energy storage device capacities, renewable energy predicted outputs, and user load time series data, etc. When the data is missing or contains noise, use the federated learning method to jointly perform distributed data cleaning and imputation with multi-region edge nodes.

[0042] Step 2: Align historical load data with meteorological factor time series based on the dynamic time warping algorithm, and realize regional load forecasting by constructing a power load spatio-temporal coupling forecasting model. This model uses a deep residual network enhanced by an attention mechanism to extract load fluctuation features, and at the same time fuses an LSTM network to capture temporal correlations.

[0043] Step 3: Construct a multi-energy coupling dispatching model with thermal power, wind power, photovoltaic power, and energy storage as decision-making units. Use a double-layer graph convolutional network to model the power flow constraints between energy nodes, define the node balance equation as the tensor product form of the topological adjacency matrix and the energy conversion efficiency, and introduce virtual nodes to handle the energy transmission losses across voltage levels.

[0044] Step 4: Establish a demand response game model based on real-time electricity price fluctuation signals, and regard the user load adjustment behavior as a non-cooperative game process under incomplete information. Use the quantum particle swarm optimization algorithm to solve the Nash equilibrium point, and then generate a time-of-use electricity price incentive strategy.

[0045] Step 5: Construct a multi-time scale rolling optimization framework and divide the dispatch cycle into three stages: day-ahead, intraday, and real-time. Use the deep reinforcement learning algorithm to build a dynamic decision-making strategy, define the state space as the grid operation parameters, the action space as the unit start and stop instructions and the energy storage charge and discharge rate, and update the policy network parameters through the double-delay deep deterministic policy gradient algorithm.

[0046] Step 6: Perform multi-objective sensitivity analysis on the scheduling parameters and use the improved Hilbert-Schmidt independence criterion to screen key variables. Construct a proxy model based on sparse polynomial chaos expansion to replace the high-dimensional optimization problem, and improve the model accuracy through adaptive sparse grid sampling.

[0047] Step 7: Design a hierarchical collaborative optimization mechanism. The outer layer uses mixed integer programming to optimize the unit combination and network topology. The inner layer uses the alternating direction multiplier method to solve the power allocation problem. The Lagrange multiplier is used to transfer boundary conditions to achieve iterative convergence of cross-layer coupling constraints.

[0048] Step 8: Construct a robust optimization model for the power flow uncertainty caused by the access of distributed power sources. Use interval fuzzy sets to describe the fluctuation range of wind and solar power output, define the robustness index as the system stability threshold under the worst scenario, and decouple the main problem from the sub-problems based on the Benders decomposition algorithm.

[0049] Step 9: Integrate the scheduling instruction verification module and use smart contract technology to perform topological security verification and N-1 fault rehearsal on the optimization results. Generate the scheduling instruction execution certificate through formal verification methods, and realize cross-domain instruction synchronization under privacy protection based on zero-knowledge proof.

[0050] Step 10: Deploy an online incremental learning mechanism and use an online sequence extreme learning machine to update the load forecasting model parameters in real time. Combine the sliding window technology to dynamically adjust the optimization model weight coefficients, and evaluate the model drift based on the Markov chain Monte Carlo method to trigger the retraining conditions.

[0051] The implementation of the present invention is further described below in conjunction with Examples 1 to 5.

[0052] Embodiment 1:

[0053] In step 1, data acquisition and processing, when there is missing or noisy data, the federated learning method is used to combine multi-region edge nodes for distributed data cleaning and interpolation.

[0054] Federated learning adopts a dynamic weight aggregation strategy. Suppose there are n edge nodes in total. For the i-th edge node, its data quality evaluation value Q i Data Integrity I i , time series continuity C iand the KL divergence of the distribution from the central node jointly determine and are obtained through weighted calculation:

[0055]

[0056] where w1, w2, and w3 are weight coefficients, and w1 + w2 + w3 = 1. These three weight coefficients can be set according to the actual data situation and experience. For example, if the integrity of the current data has a greater impact on the overall data quality, the value of w1 can be appropriately increased.

[0057] Aggregation weight ω i is calculated according to the data quality evaluation value, and the formula is:

[0058]

[0059] During the gradient transmission process, to prevent the leakage of the original data, a gradient transmission protocol based on Paillier homomorphic encryption is designed. Assume that the gradient calculated by the edge node i is g i , first use the Paillier encryption algorithm to encrypt the gradient. The public key of the Paillier encryption algorithm is (N, g), the private key is λ, and the encryption function is E. Then the encrypted gradient is E(g i ).

[0060] After the central node receives the encrypted gradients {E(g1), E(g2), …, E(g n )} from each edge node, it performs an aggregation operation. Due to the homomorphic property of the Paillier encryption algorithm, the aggregation operation can be performed in the ciphertext space, that is, calculate Finally, the central node uses the private key λ to decrypt the aggregated ciphertext to obtain the aggregated gradient for model update.

[0061] Through the above dynamic weight aggregation strategy and encrypted gradient transmission protocol, it not only fully considers the data quality differences of each edge node, improves the accuracy of data aggregation, but also effectively prevents the leakage of the original data during the transmission process, ensuring data privacy and security.

[0062] Example 2:

[0063] When constructing the spatio-temporal coupling prediction model of power load in step 2, the deep residual network plays a key role.

[0064] The deep residual network contains a residual block stacking structure. Each residual block consists of a gated convolutional layer, a batch normalization layer, and a skip connection. Let the input feature map be x, and the convolutional operation of the gated convolutional layer performs convolutional calculation on the input feature map through the convolutional kernel W to obtain the convolutional result y conv :

[0065] y conv = W * x

[0066] Among them, * represents the convolution operation.

[0067] The activation function of the gated convolutional layer adopts the Adaptive Gated Activation Function (AGAF). Let the adaptive gated activation function be σ, and its calculation formula is:

[0068]

[0069] Through the adaptive gated activation function, the gated convolutional layer can adaptively adjust information transmission according to the input features, and effectively screen out the information valuable for load fluctuation feature extraction.

[0070] After being processed by the gated convolutional layer, the data enters the batch normalization layer. The batch normalization layer normalizes the data to make it have a stable distribution, which helps to accelerate and converge the model training. Let the input of the batch normalization layer be y conv , and the output be y bn , and the calculation formula of batch normalization is:

[0071]

[0072] Among them, μ and σ 2 are the mean and variance of the current batch of data respectively, ∈ is a small constant to prevent the denominator from being zero, and γ and β are learnable parameters.

[0073] The skip connection directly adds the input x of the residual block to the output y bn after being processed by the gated convolutional layer and the batch normalization layer, to obtain the output y res of the residual block:

[0074] y res = y bn + x

[0075] This skip connection structure can effectively solve the problem of gradient disappearance in deep networks, enabling the network to learn more complex features.

[0076] Multiple residual blocks are stacked to form a deep residual network, and a self-attention module is connected to the output end of the network. Let the final output feature map of the deep residual network be F. The self-attention module first calculates the similarity between different positions in the feature map F. For each position i and j in the feature map F, calculate its similarity s ij :

[0077]

[0078] Among them, f and g are two different linear transformation functions, F i and Fj They respectively represent the feature vectors at positions i and j in the feature map F, and N is the total number of positions in the feature map F.

[0079] Through the above calculations, the spatial correlation weight matrix S=(s ij ) is obtained. The self-attention module performs weighted summation on the feature map F according to this weight matrix to obtain a feature representation with spatial correlation information, thereby more accurately capturing the spatial correlation relationship of the load characteristics and improving the accuracy of load prediction.

[0080] Example 3:

[0081] When establishing the demand response game model based on the real-time electricity price fluctuation signal in step 4, the quantum particle swarm optimization algorithm is used to solve the Nash equilibrium point.

[0082] In the quantum particle swarm optimization algorithm, the particle position encoding is represented by the quantum bit phase angle. Let the position of the particle be X=(x1,x2,…,x D ), where D is the dimension of the problem. For each dimension d, the position of the particle is represented by the phase angle θ d and

[0083] The quantum rotation gate update strategy is defined as a function of the phase difference between the particle's historical best and the global best. Let the phase angle of particle i at the t-th iteration be The phase angle of the historical best position of particle i is The phase angle of the global best position is The update formula of the quantum rotation gate is:

[0084]

[0085] where Δθ i,d is the adjustment amount of the phase angle, which is calculated by the following formula:

[0086]

[0087] Here, S(a) is the sign function. When a>0, S(a)=1; when a=0, S(a)=0; when a<0, S(a)=-1. k is a parameter that controls the rotation step size and can be adjusted according to the actual problem.

[0088] To achieve the collaborative search of the particle swarm, the quantum entanglement operation is introduced. Suppose the phase angles of particle i and particle j in a certain dimension d are θ i,d and θ j,d respectively. When performing the quantum entanglement operation, a random entanglement coefficient α∈[0,1] is selected, and then the phase angles of the two particles are updated:

[0089]

[0090]

[0091] Through quantum entanglement operations, information can be shared between particles, enhancing the collaborative search ability and preventing the algorithm from falling into local optimality.

[0092] The algorithm termination condition is that the population diversity is lower than the preset threshold or the maximum number of iterations is reached. The population diversity can be measured by calculating the standard deviation of the particle positions. Let the position vectors of the particle population be X1, X2, …, X N , and the calculation formula for the population diversity Diversity is:

[0093]

[0094] where is the average value of the positions of all particles in dimension d. When Diversity is less than the preset threshold τ, or the number of algorithm iterations reaches the maximum number of iterations T max , the algorithm stops iterating, outputs the optimal solution as the Nash equilibrium point, and then generates the time-of-use electricity price incentive strategy.

[0095] Example 4:

[0096] When performing multi-objective sensitivity analysis on the scheduling parameters and constructing a surrogate model in step 6, the sparse polynomial chaos expansion method is adopted.

[0097] The basis function of the sparse polynomial chaos expansion surrogate model selects the Legendre orthogonal polynomial. For a d-dimensional random variable z = (z1, z2, …, z d ), its general form of sparse polynomial chaos expansion is:

[0098]

[0099] where y(z) is the function to be approximated, a i is the expansion coefficient, Ψ i (z) is a multi-dimensional polynomial composed of Legendre orthogonal polynomials, and P is the total number of terms of the polynomial.

[0100] The adaptive forward-backward selection algorithm is used to determine the polynomial order and cross terms. In the adaptive forward selection stage, starting from an initial low-order polynomial model, each time a new term that can most reduce the model error is selected and added to the model. Let the current model be M k , and the model error is measured by the mean square error MSE k . For the candidate item j, calculate the model error after adding this item Select the item that makes the smallest and add it to the model, that is:

[0101]

[0102] In the adaptive backward selection stage, the selected terms are evaluated, and those terms that contribute less to reducing the model error are removed. The model error after removing each term is calculated. If the increase compared with the current model error MSE k is less than a preset threshold ∈, then this term is removed.

[0103] Optimize the model complexity based on the Akaike Information Criterion (AIC). The calculation formula of AIC is:

[0104]

[0105] where n is the number of samples, MSE is the mean square error of the model, and p is the number of parameters in the model (including the coefficients and orders of polynomials, etc.). During the model construction process, continuously adjust the order and cross terms of the polynomial, and select the model with the minimum AIC value as the final sparse polynomial chaos expansion surrogate model. In this way, while ensuring the model accuracy, overfitting of the model is effectively avoided, the generalization ability of the model is improved, so that it can more accurately replace the high-dimensional optimization problem and provide an efficient model support for subsequent optimization calculations.

[0106] Example 5:

[0107] When constructing a robust optimization model for the power flow uncertainty caused by the access of distributed power sources in step 8, the Benders decomposition algorithm is used to decouple and solve the master problem and the sub-problem.

[0108] The master problem of the Benders decomposition algorithm is modeled as a mixed integer quadratic programming problem. Let the decision variable be x (including discrete variables such as unit commitment and continuous variables such as power distribution), the objective function be f(x), and the constraint conditions include power balance constraints, equipment capacity constraints, etc. The general form of the master problem can be expressed as:

[0109]

[0110] s.t. Ax ≤ b

[0111] x ∈ X

[0112] where A is the constraint matrix, b is the constraint vector, and X is the feasible region of the decision variable.

[0113] The sub - problem is transformed into a semidefinite programming form to solve the extreme value of system stability. Let the decision variable of the sub - problem be y, which is associated with the main problem through the Lagrange multiplier λ. The objective function of the sub - problem is to maximize the system stability index g(y, x, λ) given the decision variable x of the main problem and the Lagrange multiplier λ. The constraints include power flow equation constraints, voltage stability constraints, etc. The sub - problem can be expressed as:

[0114]

[0115] s.t.Cy≤d(x)

[0116] y∈Y

[0117] where C is the constraint matrix of the sub - problem, d(x) is the constraint vector related to the decision variable x of the main problem, and Y is the feasible region of the decision variable of the sub - problem.

[0118] To improve the solution efficiency, an accelerated cutting - plane generation strategy is designed, and the sensitivity analysis of dual variables is used to screen out effective constraint conditions. According to the duality theory, the dual variable of the sub - problem reflects the influence degree of the constraint condition on the objective function. Calculate the sensitivity index of the dual variable to the decision variable of the main problem. Let the dual variable be μ, and the sensitivity index S ij is defined as:

[0119]

[0120] where i represents the number of the sub - problem constraint, and j represents the number of the decision variable of the main problem. By comparing the magnitudes of the sensitivity indices, the constraint conditions that have a greater impact on the objective function of the main problem are screened out, and only these effective constraint conditions are added to the main problem, thereby reducing the scale of the main problem and accelerating the convergence speed of the algorithm.

[0121] In step 9, the integrated scheduling instruction verification module uses a formal verification method to verify the compliance of scheduling instructions. The formal verification method uses temporal logic formulas to describe the grid security constraints. For example, the bus voltage over - limit constraint can be expressed as:

[0122]

[0123] where V(t) represents the bus voltage at time t, V min and V max are the lower and upper limits of the bus voltage respectively. The line overload constraint can be expressed as:

[0124]

[0125] where I(t) represents the line current at time t, and I max is the maximum allowable current of the line. The frequency deviation index constraint can be expressed as:

[0126]

[0127] Among them, f(t) represents the grid frequency at time t, f0 is the rated frequency, and Δf is the allowable frequency deviation range.

[0128] The compliance of the dispatching instruction is verified by traversing the state transition path through symbolic model checking technology. Symbolic model checking represents the state space of the system as a Boolean formula and uses efficient Boolean operations to verify whether the system satisfies a given temporal logic formula. During the verification process, starting from the initial state, according to the state transition relationship of the system, all possible state transition paths are gradually explored to check whether there is a violation of the grid security constraints. If the constraint conditions are satisfied in all reachable states, a dispatching instruction execution certificate is generated, indicating that the dispatching instruction is compliant; otherwise, it means that there is a risk in the dispatching instruction and it needs to be readjusted.

[0129] In step 10 of deploying the online incremental learning mechanism, an online sequential extreme learning machine is used to update the load forecasting model parameters in real time. The hidden layer node dynamic expansion strategy of the online sequential extreme learning machine adopts the error sensitivity criterion, and the node addition threshold is defined as the moving average variance of the prediction residual. Let the prediction residual be e(t), and the calculation formula of the moving average variance MSE ma (t) is as follows:

[0130]

[0131] Among them, w is the size of the moving window. When MSE ma (t) is greater than the preset node addition threshold δ, the hidden layer node expansion operation is triggered to add new nodes to improve the fitting ability of the model.

[0132] At the same time, a forgetting factor mechanism is designed to attenuate the influence weight of historical data on model update. Let the forgetting factor be β∈(0,1), and during the model update process, the weight of historical data is adjusted. For the training data at time t, its weight is β t-i , where i is the data acquisition time. In this way, as time goes by, the influence of new data on model update gradually increases, while the influence of old data gradually decreases, enabling the model to better adapt to the changes in the grid operation state.

[0133] The intelligent contract verification module includes a topology connectivity verification sub-module, which uses the depth-first search algorithm to detect the risk of island operation. Starting from a certain node in the power grid, the visited nodes are marked, and the nodes adjacent to the current node and not yet visited are recursively visited. If there are unvisited nodes after the search ends, it indicates that there is a risk of island operation in the power grid. The minimum breakpoint set is calculated based on the improved Floyd-Warshall shortest path algorithm to ensure the satisfaction rate of the N-1 criterion. The Floyd-Warshall algorithm is used to calculate the shortest path between any two points in the graph. The improvement lies in recording the path information during the calculation process. By analyzing the changes in the shortest path under different node failure conditions, the minimum breakpoint set is determined. The minimum breakpoint set refers to the smallest set of nodes that can keep the power grid connected under N-1 fault conditions after removing these nodes. During the process of calculating the minimum breakpoint set, different node combinations are continuously tried to evaluate the connectivity of the power grid, and the minimum breakpoint set that meets the N-1 criterion is found, thus ensuring the reliability of the power grid when a single component fails.

[0134] When evaluating the model drift amount based on the Markov chain Monte Carlo method, a sampler is used to generate the posterior distribution of the model parameters. Let the model parameters be θ, and a sample set {θ1, θ2, …, θ N} of the parameters is obtained through Markov chain Monte Carlo sampling. The drift amount index is defined as the distance of the parameter distribution. For example, the Kullback-Leibler divergence (KL divergence) is used to measure the difference between the current model parameter distribution P(θ) and the initial model parameter distribution Q(θ):

[0135]

[0136] When the drift amount D KL (P||Q) exceeds the preset threshold γ, the full model retraining process is triggered, and the model is retrained using the latest power grid operation data to ensure the accuracy and adaptability of the model.

[0137] Through the above embodiments, the intelligent power grid power dispatch optimization method involved in the present invention has been detailedly implemented and verified in each key technical link, which can effectively improve the optimization level of the intelligent power grid power dispatch and ensure the safe, stable and economic operation of the power grid.

[0138] It should be noted that, in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or apparatus comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or apparatus.

[0139] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent power grid power dispatching optimization method, characterized in that, It includes the following steps: Step 1: Obtain power grid operation data, including generator output curves, distribution network topology structures, energy storage device capacities, predicted output of renewable energy, and user load time series data. When there are missing or noisy data, the federated learning method is used to jointly perform distributed data cleaning and imputation with multi-region edge nodes; Step 2: Align historical load data and meteorological factor time series based on the dynamic time warping algorithm, construct a spatio-temporal coupling prediction model for power load, use a deep residual network enhanced by the attention mechanism to extract load fluctuation features, fuse the LSTM network to capture time series correlations, and generate regional-level load prediction results; Step 3: Construct a multi-energy coupling scheduling model, with thermal power, wind power, photovoltaic power, and energy storage as decision-making units, use a double-layer graph convolutional network to model the power flow constraints between energy nodes, define the node balance equation as the tensor product form of the topological adjacency matrix and the energy conversion efficiency, and introduce virtual nodes to handle cross-voltage-level energy transmission losses; Step 4: Establish a demand response game model based on real-time electricity price fluctuation signals, model the user load adjustment behavior as a non-cooperative game process under incomplete information, use the quantum particle swarm optimization algorithm to solve the Nash equilibrium point, and generate time-of-use electricity price incentive strategies; Step 5: Establish a multi-time scale rolling optimization framework, decompose the scheduling period into three stages: day-ahead, intra-day, and real-time, use the deep reinforcement learning algorithm to construct a dynamic decision-making strategy, define the state space as power grid operation parameters, the action space as unit start-stop instructions and energy storage charge-discharge rates, and update the policy network parameters through the double delayed deep deterministic policy gradient algorithm; Step 6: Conduct multi-objective sensitivity analysis on scheduling parameters, use the improved Hilbert-Schmidt independence criterion to screen key variables, construct a surrogate model based on sparse polynomial chaos expansion to replace high-dimensional optimization problems, and use adaptive sparse grid sampling to improve model accuracy; Step 7: Design a hierarchical collaborative optimization mechanism, use mixed integer programming in the outer layer to optimize unit commitment and network topology, use the alternating direction method of multipliers in the inner layer to solve the power allocation problem, and achieve iterative convergence of cross-layer coupling constraints by transmitting boundary conditions through Lagrange multipliers.

2. The intelligent power grid power dispatching optimization method according to claim 1, wherein The federated learning adopts a dynamic weight aggregation strategy, calculates the aggregation weight according to the edge node data quality evaluation value, and the data quality evaluation indicators include data integrity, time series continuity, and KL divergence from the central node distribution, and designs a gradient transmission protocol based on Paillier homomorphic encryption to prevent the leakage of original data.

3. The intelligent power grid power dispatching optimization method according to claim 1, characterized in that The deep residual network includes a residual block stacking structure, and each residual block consists of a gated convolutional layer, a batch normalization layer, and a skip connection. The activation function of the gated convolutional layer adopts an adaptive gated activation function, and a self-attention module is connected to the network output end to calculate the spatial correlation weight matrix of load features.

4. The intelligent power grid power dispatching optimization method according to claim 1, characterized in that, In the quantum particle swarm optimization algorithm, the particle position encoding is represented by the phase angle of quantum bits, the quantum rotation gate update strategy is defined as a phase difference function between the particle historical optimum and the global optimum, and a quantum entanglement operation is introduced to achieve collaborative search of the particle swarm. The algorithm termination condition is that the population diversity is lower than the preset threshold or the maximum number of iterations is reached.

5. The intelligent power grid power dispatching optimization method according to claim 1, characterized in that, The basis function of the sparse polynomial chaos expansion agent model selects Legendre orthogonal polynomials, adopts an adaptive forward-backward selection algorithm to determine the polynomial order and cross terms, and optimizes the model complexity based on the Akaike information criterion.

6. The intelligent power grid power dispatching optimization method according to claim 1, wherein The method further comprises: Aiming at the power flow uncertainty caused by the access of distributed power sources, a robust optimization model is constructed, interval fuzzy sets are used to describe the fluctuation range of wind and solar power output, the robustness index is defined as the system stability threshold under the worst scenario, and the main problem and sub-problems are decoupled and solved based on the Benders decomposition algorithm; The main problem of the Benders decomposition algorithm is modeled as a mixed integer quadratic programming problem. The subproblems are transformed into semidefinite programming forms to solve the extreme value of system stability, and an accelerated cutting plane generation strategy is designed. The dual variable sensitivity analysis is used to screen effective constraints.

7. The intelligent power grid power dispatching optimization method according to claim 1, characterized in that The method further comprises: The integrated scheduling instruction verification module uses smart contract technology to perform topological security verification and N-1 fault rehearsal on the optimization results, generates scheduling instruction execution certificates through formal verification methods, and implements cross-domain instruction synchronization under privacy protection based on zero-knowledge proof; The formal verification method adopts sequential logic formulas to describe the safety constraints of the power grid, including bus voltage over-limit, line overload and frequency deviation indicators, and verifies the compliance of the dispatching instructions by traversing the state transition path through the symbolic model detection technology.

8. The intelligent power grid power dispatching optimization method according to claim 1, characterized in that The method further comprises: An online incremental learning mechanism is deployed, and the online sequence extreme learning machine is used to update the parameters of the power load spatiotemporal coupling prediction model in real time. The sliding window technology is combined to dynamically adjust the model weight coefficients, and the Markov chain Monte Carlo method is used to evaluate the model drift to trigger the retraining conditions.

9. The intelligent power grid power dispatching optimization method according to claim 8, characterized in that, The hidden layer node dynamic expansion strategy of the online sequence extreme learning machine adopts the error sensitivity criterion, defines the node addition threshold as the moving average variance of the prediction residual, and designs a forgetting factor mechanism to attenuate the influence weight of historical data on model updating.

10. The intelligent power grid power dispatching optimization method according to claim 9, characterized in that, The Markov Chain Monte Carlo method uses a sampler to generate the posterior distribution of model parameters, defines the drift indicator as the distance of the parameter distribution, and triggers the full model retraining process when the drift exceeds a preset threshold.

Images