Power system simulation optimization method based on data driving

Abstract

The invention belongs to the technical field of power system simulation application, and particularly relates to a power system simulation optimization method based on data driving. Based on a data-driven dynamic adaptive space-time segmentation and cooperation model, a fast-varying electromagnetic transient process is mapped to a quantum processor, a slow-varying electromechanical transient process is reserved in a classical computer, a power grid node voltage is coded into quantum bit amplitude, a power grid is dynamically equivalent to quantum Hamiltonian, and a quantum Hamiltonian is obtained. And decoupling the quantum domain of the fast-changing electromagnetic transient state and the classical domain of the slow-changing electromechanical dynamic state to realize self-adaptive balance simulation precision and efficiency. Quantum calculation sparse codes and dynamic error correction are constructed, non-uniform code distance distribution is achieved, cross-scale data transmission errors are reduced, high-frequency harmonic distortion is avoided, high fault-tolerant simulation is achieved under limited quantum resources, code distances are dynamically adjusted through reinforcement learning of a DDPG strategy, noise fluctuation and power grid working condition changes are responded in real time, and resources and fault-tolerant balance are optimized.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system simulation applications, and particularly relates to a data-driven power system simulation optimization method. Background Art

[0002] In recent years, with the large-scale access of renewable energy to the power system, an increasing number of power electronic devices have brought new challenges to the stability and security of the power system. Power electronic devices contain a large number of high-frequency switching devices. The interaction between high-frequency power electronic circuits and control systems makes the dynamic characteristics of the power system become increasingly complex, and also puts forward higher requirements for power system simulation. Traditional simulation tools face problems such as a decline in accuracy and limited modeling fidelity in the simulation scenarios of power systems with power converters connected. Traditional algorithms do not have good numerical accuracy and stability, and face problems such as high computational complexity and difficulty in capturing high-frequency harmonics. Summary of the Invention

[0003] In view of the technical problems existing in power system simulation, the present invention proposes a data-driven power system simulation optimization method that is reasonable in design, simple in method, strong in theory, and can achieve the accuracy and consistency of power system simulation.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is: a data-driven power system simulation optimization method, characterized by comprising the following steps:

[0005] S1. Construct a data-driven dynamic adaptive spatio-temporal segmentation and collaboration model, map the fast-changing electromagnetic transient process to a quantum processor, retain the slow-changing electromechanical transient process in a classical computer, encode the grid node voltage as the amplitude of a quantum bit, dynamically equivalent the power network to a quantum Hamiltonian, decouple the quantum domain of the fast-changing electromagnetic transient from the classical domain of the slow-changing electromechanical dynamics, and satisfy:

[0006]

[0007] where, Δt q (t) is the step size of the quantum domain at time t, Δt c (t) is the step size of the classical domain at time t, N is a proportionality coefficient, ε is the error tolerance limit, e(t) is the local truncation error, p is the order of the implicit trapezoidal method, the boundary condition of the quantum domain predicts the switching transient through LSTM, and predicts the final interface voltage by compensating the interpolation error through CNN;

[0008] S2. Construct quantum computing sparse codes and dynamic error correction, with a non-uniform code distance distribution, divide the quantum processor into a critical area CZ and a non-critical area NCZ, implement a non-uniform code distance distribution, and the calculation formula for the sparse logic error rate is:

[0009]

[0010] Among them, p phys is the physical fault tolerance rate, d CZ is the code distance of the critical area, d NCZ is the code distance of the non-critical area, p th = 1%, C CZ is the error correction ability coefficient of the critical area, C NCZ is the error correction ability coefficient of the non-critical area, C CZ <C NCZ , and then dynamically adjust the code distance through the reinforcement learning DDPG strategy;

[0011] S3. Model implementation and operation. First, mark the converter node as the critical area in the power grid topology, and the others as non-critical areas. Set the initial code distance of the critical area to 5 and the initial code distance of the non-critical area to 3. Offline train the LSTM, CNN, and DDPG strategies, and realize the data-driven power system simulation optimization method by dynamically adaptive spatio-temporal segmentation and collaborative model assignment of computing tasks.

[0012] Preferably, the input for predicting the switch transient through LSTM is the historical voltage and switch state, and the switch transient waveform is obtained. The input for predicting the final interface voltage by compensating the interpolation error through CNN is the interpolation voltage, quantum domain current, and voltage gradient, and the voltage correction amount is obtained. The final interface voltage is:

[0013] V final = V interp +ΔV

[0014] Among them, V final is the final interface voltage, V interp is the interpolation voltage, and ΔV is the voltage correction amount.

[0015] Preferably, the code distance is dynamically adjusted through the reinforcement learning DDPG strategy. The state space is the quantum fidelity and the voltage change rate, the action space is the combination of the code distance of the critical area and the code distance of the non-critical area, and the formula of the reward function is:

[0016] R(t) = w1Accuracy - w2Latency - w3Qubitcost

[0017] Among them, w1, w2, and w3 are weight coefficients, Accuracy represents the reciprocal of the error between the simulation result and the voltage amplitude of the real physical system, Latency represents the time delay from the start of the simulation task to the output of the result, and Qubitcost represents the total number of physical qubits consumed during the simulation process.

[0018] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0019] The present invention designs a data-driven simulation optimization method for power systems, constructs a data-driven dynamic adaptive spatio-temporal segmentation and cooperation model, maps the fast-changing electromagnetic transient process to a quantum processor, retains the slow-changing electromechanical transient process in a classical computer, encodes the grid node voltage as the amplitude of a quantum bit, dynamically equivalent the power network to a quantum Hamiltonian, decouples the quantum domain of the fast-changing electromagnetic transient and the classical domain of the slow-changing electromechanical dynamics, and realizes the adaptive balance between simulation accuracy and efficiency.

[0020] Construct quantum computing sparse codes and dynamic error correction, with non-uniform code distance distribution, reduce cross-scale data transfer errors, avoid high-frequency harmonic distortion, and achieve high-fault-tolerant simulation under limited quantum resources. Dynamically adjust the code distance in real time through the DDPG strategy of reinforcement learning to respond to noise fluctuations and grid operating conditions changes, and optimize the balance between resources and fault tolerance. Description of the Drawings

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0022] Figure 1 It is a schematic structural diagram of a data-driven simulation optimization method for power systems. Detailed Embodiments

[0023] In order to better understand the above-mentioned objects, features, and advantages of the present invention, the following will further illustrate the present invention with reference to the drawings and embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0024] Many specific details are set forth in the following description to facilitate a thorough understanding of the present invention. However, the present invention may be implemented in other ways different from those described herein. Therefore, the present invention is not limited by the specific embodiments disclosed in this specification.

[0025] Embodiment, as Figure 1 As shown, considering that power electronic devices contain a large number of high-frequency switching devices, the interaction between high-frequency power electronic circuits and control systems makes the dynamic characteristics of power systems become increasingly complex, and also puts forward higher requirements for power system simulation. Traditional simulation tools face problems such as reduced accuracy and limited modeling fidelity in the simulation scenarios of power systems with power converters connected. Traditional algorithms do not have good numerical accuracy and stability, and face problems such as high computational complexity and difficulty in capturing high-frequency harmonics. The present invention proposes a data-driven simulation optimization method for power systems.

[0026] As the power system continues to develop, its complexity and data size are increasing. Traditional computing methods and technologies can no longer meet the power system's needs for efficient, accurate computing and decision-making. As a new computing model, quantum computing has great application potential in the power system. As the power system continues to develop, its complexity and data size are increasing. Traditional computing methods and technologies can no longer meet the power system's needs for efficient, accurate computing and decision-making. As a new computing model, quantum computing has great application potential in the power system. In terms of power grid planning and optimization, quantum computing can efficiently solve complex optimization problems.

[0027] LSTM network is a special type of recurrent neural network, which has the advantages of high prediction accuracy and strong recursive ability, and can solve the problem of long-term dependence of recurrent neural networks. Adding a "gate" structure to add or discard information can play a role in long-term memory.

[0028] Convolutional Neural Network (CNN) has been widely used in image classification, target detection and other fields due to its excellent feature extraction ability and good parallelism. CNN contains a large number of repeated multiplication and accumulation operations. Reconfigurable processors have shown significant advantages in CNN computing research due to their regular parallel structure and flexible configuration capabilities.

[0029] Considering the superposition and entanglement characteristics of quantum states, large-scale linear equations can be solved in parallel, the complexity is reduced, and quantum evolution follows unitary transformation. In theory, the energy consumption thermodynamic limit of classical computers can be avoided, and the electromagnetic transient (fast-changing) process is mapped to the quantum processor, and the electromechanical transient (slow-changing) process is retained in the classical computer to achieve time-space separation, calculate boundary conditions, and achieve time-space dimension decoupling. Therefore, a data-driven dynamic adaptive time-space segmentation and coordination model is first constructed to map the fast-changing electromagnetic transient process to the quantum processor, and the slow-changing electromechanical transient process is retained in the classical computer. The grid node voltage is encoded as the quantum bit amplitude, and the power network dynamics are equivalent to the quantum Hamiltonian, decoupling the quantum domain of fast-changing electromagnetic transients and the classical domain of slow-changing electromechanical dynamics, satisfying:

[0030]

[0031] Among them, Δt q (t) is the quantum domain step size at time t, Δt c(t) is the classical domain step size at time t, N is the proportionality coefficient, ε is the error tolerance limit, e(t) is the local truncation error, p is the order of the implicit trapezoidal method, which realizes the relationship between the adaptive balance of the simulation progress and efficiency, dynamically adjusts the step size according to the local error, automatically reduces the step size (in the order of μs) during the high-frequency switching stage of the converter (large error), and amplifies the step size (in the order of ms) during the low-frequency dynamic stage of the power grid (small error), achieving the balance between accuracy and efficiency. Compared with the fixed-step method, the overall computational amount is reduced, which is suitable for long-time simulation scenarios. The quantum domain boundary conditions predict the switching transient through LSTM and predict the final interface voltage by compensating the interpolation error through CNN. Specifically, the input for predicting the switching transient through LSTM is the historical voltage and the switching state, obtaining the switching transient waveform. The input for predicting the final interface voltage by compensating the interpolation error through CNN is the interpolated voltage, the quantum domain current, and the voltage gradient, obtaining the voltage correction amount. The final interface voltage is:

[0032] V final =V interp +ΔV

[0033] where, V final is the final interface voltage, V interp is the interpolated voltage, and αV is the voltage correction amount. LSTM captures the temporal characteristics of the switching event, and CNN corrects the local non-linear error, reducing the relative error at the μs-level mutation point of the interface voltage, integrating the quantum domain current and the classical domain gradient information, eliminating the phase lag problem of traditional interpolation methods (such as cubic spline), and improving the harmonic analysis accuracy, achieving cross-scale consistency.

[0034] Deep Reinforcement Learning (DRL) is one of the core technologies of artificial intelligence, which provides an efficient method for decision-making in various fields by the interaction and learning between the agent and the environment. To solve the problem of the continuous action space, the Deterministic Policy Gradient (DPG) algorithm is proposed. Combining the ideas of the Deep Q-Network (DQN) and DPG, the Deep Deterministic Policy Gradient (DDPG) algorithm is created.

[0035] Then, considering that the traditional surface code needs to arrange qubits into a dense two-dimensional grid, the resource overhead is huge. According to the characteristics of power system simulation, the converter nodes (high-frequency switches) require high fidelity, while the transmission lines (low-frequency dynamics) can tolerate higher noise. The error rate of the converter switches increases suddenly during the switching instant and is relatively low at steady state. Constructing a quantum computing sparse code and dynamic error correction, with a non-uniform code distance distribution, dividing the quantum processor into a critical area CZ and a non-critical area NCZ, implementing a non-uniform code distance distribution, the calculation formula for the sparse logical error rate is:

[0036]

[0037] Among them, p phys is the physical fault tolerance rate, d CZ is the code distance of the critical area, d NCZ is the code distance of the non-critical area, p th = 1%, C CZ is the error correction ability coefficient of the critical area, C NCZ is the error correction ability coefficient of the non-critical area, C CZ < C NCA , and then dynamically adjust the code distance through the reinforcement learning DDPG strategy. The non-uniform code distance design reduces the consumption of physical qubits and allows dynamic adjustment of the CZ / NCZ partition. In response to changes in the power grid topology (such as the access of new energy), there is no need to reconstruct the entire error correction code. Specifically, the dynamic adjustment of the code distance through the reinforcement learning DDPG strategy has a state space of quantum fidelity and voltage change rate, an action space of the combination of the code distance of the critical area and the code distance of the non-critical area, and the formula of the reward function is:

[0038] R(t) = w1Accuracy - w2Latency - w3Qubitcost

[0039] Among them, w1, w2, and w3 are weight coefficients. Accuracy represents the reciprocal of the error between the simulation result and the voltage amplitude of the real physical system. Latency represents the time delay from the start of the simulation task to the output result. Qubitcost represents the total number of physical qubits consumed during the simulation process. The weight coefficients seek the optimal between accuracy, speed, and resources, improve the comprehensive efficiency, and support dynamic adjustment of weights such as during faults and adapt to different operating modes at steady state. By penalizing the qubit cost, it avoids over-occupying resources on medium-scale quantum devices with noise and improves the task parallelization ability.

[0040] Finally, the model is implemented and run. In the power grid topology, the converter nodes are first marked as the critical area, and the others are the non-critical area. The initial code distance of the critical area is set to 5, and the initial code distance of the non-critical area is set to 3. The LSTM, CNN, and DDPG strategies are trained offline. By dynamically adapting spatio-temporal segmentation and collaborative model to allocate computing tasks, a data-driven power system simulation optimization method is realized.

[0041] The above is only a preferred embodiment of the present invention, and it is not a limitation of the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical content of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A data-driven power system simulation optimization method, characterized in that: The steps include: S1. Build a data-driven dynamic adaptive space-time segmentation and coordination model, map the fast-changing electromagnetic transient process to the quantum processor, keep the slow-changing electromechanical transient process in the classical computer, encode the grid node voltage as quantum bit amplitude, and dynamically equate the power network to the quantum Hamiltonian, decouple the quantum domain of fast-changing electromagnetic transients and the classical domain of slow-changing electromechanical dynamics, and satisfy: Among them, Δt q (t) is the quantum domain step size at time t, Δt c (t) is the classical domain step size at time t, N is the proportionality coefficient, ε is the error tolerance limit, e(t) is the local truncation error, p is the implicit trapezoidal method order, and the quantum domain boundary conditions are used to predict the switching transients through LSTM, and the final interface voltage is predicted through CNN to compensate for the interpolation error; S2. Construct quantum computing sparse code and dynamic error correction, non-uniform code distance distribution, divide the quantum processor into critical area CZ and non-critical area NCZ, implement non-uniform code distance distribution, and calculate the sparse logic error rate as follows: Among them, p phys is the physical fault tolerance, d CZ is the key area code distance, d NCZ is the code distance in the non-critical area, p th =1%, C CZ is the error correction capability coefficient of the critical area, C NCZ is the error correction capability coefficient of the non-critical area, C CZ <C NCZ , and then dynamically adjust the code distance through reinforcement learning DDPG strategy; S3. Implement data-driven power system simulation optimization. In the power grid topology, first mark the converter node as the critical area and the others as non-critical areas. Set the initial critical area code distance to 5 and the initial non-critical area code distance to 3. Train LSTM, CNN and DDPG strategies offline. Allocate computing tasks through dynamic adaptive spatiotemporal segmentation and collaborative models to implement data-driven power system simulation optimization methods.

2. The data-driven power system simulation optimization method according to claim 1, characterized in that: The input of the switch transient prediction by LSTM is the historical voltage and switch state, and the switch transient waveform is obtained. The input of the final interface voltage prediction by CNN compensation interpolation error is the interpolation voltage, quantum domain current and voltage gradient, and the voltage correction is obtained. The final interface voltage is: V final =V interp +ΔV Among them, V final is the final interface voltage, V interp is the interpolation voltage, and ΔV is the voltage correction.

3. The data-driven power system simulation optimization method according to claim 1 is characterized in that: The code distance is dynamically adjusted through the reinforcement learning DDPG strategy. The state space is the quantum fidelity and the voltage change rate. The action space is the combination of the critical area code distance and the non-critical area code distance. The formula of the reward function is: R(t)=w1Accuracy-w2Latency-w3Qubitcost Among them, w1, w2 and w3 are weight coefficients, Accuracy represents the inverse of the error between the simulation result and the voltage amplitude of the real physical system, Latency represents the time delay from the start of the simulation task to the output of the result, and Qubitcost represents the total number of physical quantum bits consumed during the simulation process.

Images