Electric ship charging load scene generation method based on D-W-CGAN and computer system

By introducing non-Markov process and conditional generation adversarial networks in the generation of electric ship charging load scenarios, the problem of difficulty in simulating electric ship charging load scenarios in the prior art is solved, efficient scene generation and feature mining are achieved, and the model is improved stability and accuracy.

CN118965149APending Publication Date: 2024-11-15STATE GRID FUJIAN POWER ELECTRIC CO ECONOMIC RESEARCH INSTITUTE +1

Patent Information

Application Number
CN202410921386.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-10
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing technology is difficult to accurately characterize and simulate the charging load scenarios of electric ships in complex water transport environments, which leads to challenges in planning and scheduling of port power systems, especially the problems of noise interference and potential probability distribution characteristics in data training are difficult to explore.

Method used

Using the D-W-CGAN-based electric ship charging load scenario generation method, forward noise is performed by introducing a non-Markov process, and reverse noise is performed by combining the condition generation adversarial network to generate massive charging load scenarios, using diffusion model and generating confrontation Network mining potential probability distribution characteristics.

Benefits of technology

Improve the forward noise addition efficiency and reverse noise denoising capability of the diffusion model, generate high-quality charging load scenarios, solve the instability problem in GAN training, enhance the robustness and generalization capabilities of the model, and can effectively Simulate the characteristics and scenarios of the charging load of the ship.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118965149A_ABST
    Figure CN118965149A_ABST
Patent Text Reader

Abstract

The invention relates to a D-W-CGAN-based electric ship charging load scene generation method and a computer system. The method comprises the following steps: constructing a training set containing charging load time sequence data of each charging pile and corresponding environment data; gaussian noise is added into the training set through the forward diffusion process of the diffusion model to reconstruct the training set so as to generate noise data; training a conditional generative adversarial network: inputting the noise data and the random noise into a generator of the conditional generative adversarial network, so that the generator establishes a denoising function capable of removing Gaussian noise in the noise data by learning a distribution rule of a training set so as to recover a latent probability distribution feature of the training set, and inputting new data generated by the training set and the generator by using a denoising function into a discriminator for distinguishing, wherein the generator and the discriminator continuously optimize own parameters in the training process. And inputting the charging load time sequence data and the random noise into the trained model to generate a charging load scene.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of power system and artificial intelligence technology, and specifically relates to a method and a computer system for generating charging load scenarios of electric ships based on DW-CGAN. Background Art

[0002] Due to the low efficiency and serious emissions of traditional ships, electric ships (ES) have gradually become a hot topic in the shipping field with their advantages of zero emissions, high efficiency, low cost, and clean and low carbon. However, electric ships have long sailing time and short port stays, so charging stations usually need to be equipped with ultra-high power fast charging, which will lead to an increasing contradiction between power supply and demand in ports. As the ship charging load increases, its strong randomness will have a great impact on the three-phase load balance of the distribution network, the life of the distribution transformer, the local overload of the distribution network load, and the power quality of the distribution network. Therefore, the future high penetration rate of port charging load connection will bring new challenges to the planning of port power systems, especially the planning and operation of charging stations. Energy management and charging facility planning and layout of inland waterway transportation have become urgent issues to be solved. How to accurately characterize the uncertainty of ship charging load is the key to meeting this challenge. As a technical means to accurately characterize the uncertainty of load output, scenario method is widely used in power system planning and scheduling operation.

[0003] However, due to limitations in data collection and technological development, no research has been conducted on the generation of ship charging load scenarios. In addition, the travel mode of electric ships contains multi-dimensional and complex characteristics, which are easily affected by noise. Data training faces certain difficulties, especially when the data information is relatively complex. It is difficult to effectively simulate its potential probability distribution characteristics using existing methods. Therefore, it is difficult for existing methods to mine the characteristics of ship charging loads and generate charging scenarios that meet the needs of electric ships in complex water transport environments. Summary of the invention

[0004] Aiming at the problem of lack of electric ship charging load scenario generation in complex water transport environment, the present invention proposes a method and system for generating electric ship charging load scenarios based on DW-CGAN. Firstly, in order to improve the forward noise efficiency of the diffusion model, a non-Markov process is introduced for forward noise addition; then, in order to improve the reverse denoising ability and reasoning learning speed of the diffusion model, a conditional generative adversarial network is used for implicit modeling; finally, the DW-CGAN model is used to generate a large number of scenarios of electric ship charging load.

[0005] In the first aspect, a model for generating charging load scenarios for electric ships is provided, including: a diffusion model, which uses its forward diffusion process to add Gaussian noise to the charging load time series data to reconstruct it to generate noise data that conforms to the Gaussian distribution; a conditional generative adversarial network, whose generator uses a denoising function to reversely remove the Gaussian noise in the noise data to restore the potential probability distribution characteristics of the charging load time series data, thereby realizing the generation of charging load scenarios for electric ships. The charging load time series data includes the charging start time, end time, charging power, and charging location / charging pile number. The Gaussian noise is added using a non-Markov process noise addition method.

[0006] In the second aspect, a training method for generating a model for charging load scenarios of electric ships is provided, including: constructing a training set, the training set including charging load time series data of each charging pile and corresponding environmental data; adding Gaussian noise to the training set through the forward diffusion process of the diffusion model to reconstruct it to generate noise data that conforms to the Gaussian distribution; training a conditional generative adversarial network: inputting noise data and random noise into the generator of the conditional generative adversarial network, so that the generator can establish a denoising function that can remove Gaussian noise in the noise data and restore the latent probability distribution characteristics of the training set by learning the distribution law of the training set, inputting the training set and the new data generated by the generator using the denoising function into the discriminator for distinction, and during the training process, the generator and the discriminator continuously optimize their own parameters, and the training ends when the discriminator determines that the charging training set and the new data reach Nash equilibrium. The environmental data include wind speed, wind direction and temperature. During the training process, the Wasserstein distance is used to measure the gap between the distribution of the new data generated by the generator and the true distribution of the training set.

[0007] In a third aspect, a method for generating a charging load scenario for an electric ship is provided, wherein charging load time series data and random noise are input into a model trained using the training method for generating a charging load scenario model for an electric ship, and the model generates a charging load scenario. The charging load time series data is predicted by historical charging load time series data and corresponding environmental data.

[0008] In a fourth aspect, a computer system is provided, comprising: a processor; a memory, comprising one or more computer program modules; wherein the one or more computer program modules are stored in the memory and configured to be executed by the processor, and the one or more computer program modules include instructions for implementing the training method of the electric ship charging load scenario generation model and the instructions for the electric ship charging load scenario generation method.

[0009] In a fifth aspect, a computer-readable storage medium is provided for storing non-temporary computer-readable instructions, which, when executed by a computer, can implement the training method for the electric ship charging load scenario generation model and the electric ship charging load scenario generation method.

[0010] Charging utility companies can use the present invention to simulate the complex distribution of historical ship loads and generate a deterministic set of scenarios to comprehensively characterize the uncertainty of the load, so that traditional planning and scheduling models and solution methods for deterministic problems can continue to be used, effectively improving the applicability of traditional planning and scheduling solution strategies to charging load uncertainty problems.

[0011] (1) The present invention introduces a non-Markov process to process the forward process of the diffusion model. The process does not depend on the previous state, so the forward noise addition efficiency of the diffusion model can be significantly improved.

[0012] (2) In the reverse denoising process, the present invention adopts a conditional generative adversarial network for implicit modeling and inference learning. The conditional generative adversarial network predicts the next denoising data based on the current noise data, thereby achieving effective elimination of noise. Therefore, it can fully utilize the advantages of GAN, that is, through adversarial training, it can generate high-quality data.

[0013] (3) The present invention not only solves the instability problem in GAN training, but also maintains the advantages of the diffusion model, such as generating high-quality samples and stable training, while also taking into account the advantages of GAN in generating sample diversity. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 4 is a flow chart of a method for generating a charging load scenario for an electric ship according to an embodiment of the present invention.

[0015] Figure 2 It is a logic block diagram of generating an electric ship charging load scenario based on DW-CGAN according to an embodiment of the present invention.

[0016] Figure 3 Schematic diagram of the change of the discriminator loss function of DW-CGAN during the training process according to an embodiment of the present invention.

[0017] Figure 4 Schematic diagram of the Wasserstein distance between the real sample distribution and the generated sample distribution according to an embodiment of the present invention.

[0018] Figure 5a 1 is a box plot of original samples in several scenarios according to an embodiment of the present invention.

[0019] Figure 5b 1 is a box plot of samples generated by the generator of DW-CGAN according to an embodiment of the present invention in several scenarios.

[0020] Figure 6 4 is a flow chart of constructing an electric ship charging load data set according to an embodiment of the present invention.

[0021] Figure 7 : is a diagram of an electric ship charging load scenario generated by DW-CGAN according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] Figure 1 The flow chart of a method for generating charging load scenarios for electric ships based on DW-CGAN is shown. The method includes: obtaining multi-source characteristic information such as basic parameters of each electric ship, historical traffic travel patterns and meteorology; constructing an original data set based on the multi-source characteristic information; mining the potential probability distribution characteristics of the electric ship charging time load series data under multi-source characteristic information by training the forward diffusion network and the reverse diffusion network in the DW-CGAN model; generating electric ship charging load scenarios using the established DW-CGAN model; and conducting quality evaluation on the generated load scenarios in terms of accuracy and generalization.

[0023] The logic diagram of the electric ship charging load scenario generation based on DW-CGAN is as follows: Figure 2 As shown. The non-Markov process is used to positively add noise to the original data, which is both efficient and reliable. Then the conditional generative adversarial network is used for implicit modeling to achieve reverse denoising. The feature information matrix (charging load data and environmental data) is flattened into a vector and input into the generator of the conditional generative adversarial network together with the noise extracted from the normal distribution to generate the charging load matrix. By training the generator and discriminator in CGAN, the charging load scenarios of electric ships under different labels (labels divided by season) are obtained. The following describes in detail the generation method, model and model training method of the charging load scenario of electric ships.

[0024] Step 1: Obtain multi-source characteristic information such as basic parameters of each electric ship, historical traffic travel patterns, and meteorology.

[0025] Step 2: Construct the original dataset (also called: original sample / real sample) based on multi-source feature information.

[0026] Step 2.1: Calculate the energy consumption of each ship based on the ship energy consumption model.

[0027] The navigation environment of different sections in the inland river basin varies greatly. According to the environmental characteristics of different sections, the inland river basin is divided into M seasons and N sections. By calculating the fuel consumption of each section in a certain season, the fuel consumption of one voyage in that season is obtained. The power consumption of the round-trip voyage in different seasons is accumulated to obtain the annual power consumption of the ship. The annual power consumption is as follows:

[0028]

[0029] Where: Q is the total annual electricity consumption of the all-electric ship; c is the number of voyages of the all-electric ship in the nth voyage segment in season m, G m,n D is the power consumption per unit distance of the all-electric ship in the nth voyage in season m; n is the distance of the nth flight segment; M is the total number of seasons; N is the total number of flight segments; Q aux The total fuel consumption of auxiliary engines. aux =0.2Q.

[0030] Step 2.1.1: Calculate the power consumption per unit distance.

[0031] The power consumption per unit distance of the main engine of a fully electric ship in the nth voyage section in season m is as follows:

[0032]

[0033] Where: P m,n is the main engine output power of the all-electric ship in the nth segment in season m, in kW; v m,n It is the ground speed of the all-electric ship in the nth segment of season m, in km / h.

[0034] Step 2.1.2: Calculate the ship's main engine power.

[0035] When an all-electric ship sails at different speeds, the main engine of the all-electric ship will generate different powers to overcome the sailing resistance of the ship. According to the matching relationship between ship, engine and propeller, the main engine power of the all-electric ship at a certain speed can be expressed by the following formula:

[0036]

[0037] Where: P m,n is the output power of the main engine of the all-electric ship; R m,n is the sailing resistance of the all-electric ship in the nth section of season m; K is the number of propellers; η S is the shaft transmission efficiency; η G is the gearbox efficiency; η R is the relative rotation efficiency of the propeller; η H is the hull efficiency; η0 is the propeller open water efficiency; V S,m,nis the still water speed of the all-electric ship in the nth segment of season m.

[0038] Step 2.1.3: Calculate the sailing speed.

[0039] V S,m,n =v g,m,n +v w,m,n (4)

[0040] Where: v g,m,n is the ground speed of the all-electric ship in the nth segment in season m, km / h; v w,m,n is the water flow speed of the all-electric ship in the nth section in season m, km / h.

[0041] Step 2.1.4: Calculate the resistance of each ship based on the sailing resistance model.

[0042] During the voyage, the ship's hull will be subject to resistance from wind, waves, currents, etc. The energy consumption of the ship's main engine is mainly used to overcome the sailing resistance so that the ship can reach a certain speed. The sailing resistance of the ship can be expressed as:

[0043] R m,n =R T,m,n +R wave,m,n +R wind,m,n +R shallow,m,n =F(v g,m,n ,v w,m,n ,v wind,m,n ,H m,n ) (5)

[0044] Where: R m,n R is the total resistance of the ship; T,m,n is the static water resistance of the ship; R wave,m,n is the wave-making resistance of the ship; R wind,m,n is the wind resistance of the ship; R shallow,m,n is the shallow water resistance of the ship; v g,m,n v is the ship's ground speed in the nth leg of season m; w,m,n is the water velocity of the nth section in season m; v wind,m,n H is the wind speed of the nth segment in season m; m,n is the channel water depth of the nth section in season m.

[0045] Step 2.2: Calculate the charging load of each charging pile in the port

[0046] Each charging pile has its own charging event data. For each charging event, the load value per minute during the charging time is calculated according to the following formula:

[0047]

[0048] In the formula, xt It indicates the load value per minute during the charging time, in kW; T is the charging time, in minutes; Q T The electric energy allocated during the charging time is expressed in kWh. T It is converted from the total annual electricity consumption Q of all-electric ships.

[0049] Step 2.3: Establish a training set (dataset) for training the DW-CGAN model. The training set contains the charging load time series data of each charging pile and the corresponding environmental data. The charging load time series data contains the charging start time, end time, charging power and charging location (charging pile number). Environmental data can include wind speed, wind direction, and temperature. The data set has spatiotemporal characteristics. The charging load of each charging pile can be obtained by actual measurement. If the charging pile has not been arranged yet, the charging load can be calculated by the above formulas (1) to (6).

[0050] Step 3: Construct the DW-CGAN model. By training the forward diffusion network and the reverse diffusion network in the DW-CGAN model, the latent probability distribution characteristics of the electric ship charging load sequence data input under multi-source feature information are mined.

[0051] The diffusion model (Denising Diffusion Probabilistic Model, DDPM) consists of a forward process q and a reverse process p θ The forward process q gradually adds smaller Gaussian noise to the charging load time series data through T-step iterations until the data structure is completely destroyed. The reverse process p θ The added noise is gradually removed through the denoising function, and the original data is finally restored.

[0052] Step 3.1: Forward diffusion process.

[0053] In the process of scenario generation for electric ship charging load, adding Gaussian noise to the original load data can not only increase the randomness of the data, but also improve the model's tolerance to noise, thereby enhancing the robustness and generalization ability of the model. The present invention adopts a non-Markov process (Non-Markov process) noise addition method for forward noise addition. A non-Markov process refers to a general process in which the probability distribution of a random variable does not belong to a Markov process. It is different from the Markov process in that its characteristic is that the value of the current time step is only related to the state at the current moment, but has nothing to do with the value of the previous time step, that is, its transition probability depends not only on the current state, but also on the previous state. The specific steps are as follows: Define q(x0) as the charging load time series data in the real space The forward process is defined as:

[0054]

[0055]

[0056] Where: (L is the data dimension), t is the index of the diffusion step, t =0,1,…,T,β1,β2,…,β T is a constant between 0 and 1, q(x t x t-1 ) conforms to the Gaussian distribution shown in formula (8).

[0057] An important property of forward diffusion is that we can explicitly calculate x at any intermediate step based on the initial electric ship charging load time series data x0 t :

[0058]

[0059] Where: Where β s is the diffusion rate at step size s (i.e., the variance of Gaussian noise), which increases with the forward modeling process.

[0060] Then, we can deduce that x t The closed form expression for is:

[0061]

[0062] Step 3.2: Reverse diffusion process.

[0063] In order to obtain the real charging load time series data, the present invention combines non-equilibrium thermodynamics and designs a progressive inverse process to transform the simple Gaussian distribution into the actual ship charging load time series data distribution. The real charging load time series data distribution is restored by calculating the step-by-step Gaussian noise injected in the forward process and performing a series of iterative denoising steps. Therefore, the reverse process is from pure Gaussian x T Starting from the perturbation x T The Gaussian noise is removed step by step until the real charging scene is fully restored. The reverse process is defined as:

[0064]

[0065]

[0066] Where: Reverse process p θ (x t-1 |x t ) conforms to the Gaussian distribution shown in formula (12), θ is a learnable parameter, μ θ (xt,t) and are the mean and variance of the reverse process respectively. In the reverse process, the Gaussian noise x t-1 ~N(0,I) for initial sampling, and then use the Langevin dynamics sampling method to sample x t-1 ~p θ (x t-1 |x t ) is iteratively sampled to learn a suitable denoising function and finally restore the original charging load time series data distribution.

[0067] Step 3.2.1: Conditional Generative Adversarial Network Modeling Denoised Distribution.

[0068] Based on the non-Markov process, the present invention constructs a forward noise addition process of the diffusion model. In this process, assuming that the value of T is small, it means that each diffusion step has a large β t value.

[0069] In order to link this diffusion process with the conditional generative adversarial network, a metric generator distribution p is introduced θ (x t-1 |x t ) and the true denoised distribution q(x t-1 | t ), that is, by calculating the divergence between the two, as shown in formula (13), the real denoising distribution can be simulated and approximated efficiently and accurately.

[0070]

[0071] Where: is the log-likelihood ratio; represents the Kullback-Leibler (KL) divergence; Wasserstein distance is used to measure the generator distribution p θ (x t-1 |x t ) and the true denoised distribution q(x t-1 | t The network parameters are updated using the non-saturated gradient descent method.

[0072] "Generator distribution" refers to the distribution of data generated by the generator, and "true denoising distribution" refers to the distribution of real data (training set when training the model). When the Nash equilibrium is reached between the generator distribution and the true denoising distribution, that is, the generator is able to generate samples that are almost indistinguishable from the true data distribution, it means that the generator has successfully captured the exact distribution of the real data.

[0073] Step 3.2.2: Discriminator design.

[0074] The discriminator is expressed as:

[0075]

[0076] Where: are the parameters of the discriminator, x t-1 and x t Is the input of the discriminator. The input data of the discriminator is the real data and the data generated by the generator. The function of the discriminator is to distinguish whether the data comes from reality or generated simulation.

[0077] The loss function of the discriminator is established by the following formula:

[0078]

[0079] In the formula: the first term is the real data sample loss, and the second term is the generated sample loss. The real data sample loss determines the ability of the discriminator to recognize real data, and the generated sample loss determines the ability of the discriminator to recognize the data generated by the generator. Equivalent transformation of formula (15) leads to formula (16), that is:

[0080]

[0081] Step 3.2.3: Generator design.

[0082] The loss function of the generator is the objective function of the generator. The input of the generator is the actual data and random noise, and the output is the generated data obtained by learning the rules of the actual data.

[0083] Based on the given discriminator, the loss function of the generator is given as (17), namely:

[0084]

[0085] Formula (17) uses non-saturated GAN to train the generator. Specifically, parameterized implicit modeling is used for the reverse denoising process. The diffusion model does not directly predict xt-1 in the denoising step, but first establishes a parameterized denoising model pθ(xt-1|xt), that is:

[0086] p θ (x t-1 |x t ):=q(x t-1 |x t ,x0=f θ (x t ,t)) (18)

[0087] In formula (18), t is the index of the diffusion step, t = 0, 1, ..., T, x t is the noise data of step t, x t-1is the noise data of step t-1, firstly through the implicit denoising model f θ (x t ,t) to predict x0, and then given x t and predicted value x0, through the posterior distribution q(x t-1 |x t ,x0) to sample x t-1 Intuitively speaking, q(x t-1 |x t ,x0) is x t-1 The distribution on x t When x0 is de-noised, it always has the form of a Gaussian distribution. θ (x t-1 |x t ), then equation (18) can be further written as follows:

[0088]

[0089] In the formula: z is a hidden variable, z~p(z):=N(z,0,I).

[0090] Step 3.3: Use the normalized training data set (including the historical charging load time series data of each charging pile and the corresponding environmental data) to train the DW-CGAN model. When the training is completed, extract the final network parameters of the model.

[0091] During model training, when the discriminator converges to similar output values ​​of D(G(z)) and D(x), the generator can generate real ship charging load samples. The discriminator loss function of DW-CGAN changes during the training process as shown in the figure below: Figure 3 As shown in the figure, the two curves represent the evolution of the output generated sample G(z) of the discriminator D(x) / D(G(z)) during the training of the dataset. It can be seen that at the beginning, the generated samples (dark curve in the figure) and the actual samples (light curve in the figure) are easily distinguished on the discriminator. As the training progresses, they get closer and closer.

[0092] Figure 4 It represents the Wasserstein distance between the real sample distribution and the generated sample distribution, where being close to zero means that the two distributions are close to each other. As can be seen from the figure, when the number of training times reaches 7852, the network basically tends to a stable state, and both the generator and the discriminator networks reach convergence.

[0093] Figure 5a , Figure 5bBox plots of the original samples and generated samples in several scenarios are given respectively, indicating that the method proposed in the present invention is closest to the actual data of ship charging load and can effectively restore the distribution law and time series distribution of the actual data of electric ship charging load.

[0094] Step 4: Use the DW-CGAN model established in step 3 to generate the electric ship charging load scenario.

[0095] Step 3 is to use historical charging load data and environmental data to train the DW-CGAN model. Step 4 can use the DW-CGAN model trained in step 3 to generate charging load scenarios for other time periods.

[0096] If you want to generate future charging load scenarios, you first need to build a charging load dataset. Figure 6 As shown in the figure, the specific process is as follows: collect the historical load data of the charging pile and the corresponding environmental data, including the charging start time, end time, charging power, and charging location (charging pile number); establish an algorithm model that can realize spatiotemporal prediction (this algorithm model is a conventional technology in this field); input the historical load data of the charging pile and the corresponding environmental data into the algorithm model, and use the algorithm model to predict the future charging load of the charging pile. Sum the predicted load sequence of the charging pile by hour, and finally obtain the load time series of the charging pile. Then do the same processing for all charging events of the port charging pile to obtain the charging load time series data.

[0097] The predicted charging load time series data and random noise are input into the DW-CGAN model to obtain the electric ship charging load scenario set. Figure 7 The generation results of a charging load time series are shown, and each curve represents a typical class of generated charging demand samples, which have similar time patterns.

[0098] Step 5: Generate quality assessment of load scenarios, including accuracy and generalization.

[0099] Step 5.1: Accuracy evaluation. The coverage and accuracy of the generated charging load scenarios are measured, and the accuracy is confirmed using the root mean square error (RMSE) and mean square error (MSE).

[0100] RMSE is the root mean square of the ratio of the square of the deviation between the actual sample and the generated sample to the number of samples N, as shown in formula (20). RMSE can reflect the deviation between the actual sample and the generated sample in scene generation. The smaller the value, the higher the accuracy of the generated sample.

[0101]

[0102] Where: true value yi is the data value of the actual sample; is the data value for generating samples; N is the number of samples.

[0103] MSE is the average value of the sum of the absolute values ​​of the differences between the actual samples and the generated samples, as shown in formula (21). MSE can only measure the average modulus of the generated sample errors. The larger the value, the greater the error between the generated samples and the actual samples.

[0104]

[0105] Step 5.2: Generalization evaluation: The generalization is confirmed by measuring the fit between the generated charging load data samples and the actual samples, i.e. R-squared (R2).

[0106] R 2 It describes the degree of fit between the generated sample and the actual sample, as shown in formula (22). 2 It can intuitively show the degree of closeness between the generated scene and the actual data sample. The closer its value is to 1, the closer the data of the generated scene is to the real value.

[0107]

[0108] Where: for y i The average value of .

[0109] In order to further illustrate the superiority of the proposed method, the ship charging load scenarios generated by the three methods of WGAN, AC-GAN, and RAC-GAN are compared with the algorithm proposed in this invention. The evaluation indicators of the charging load generated by the above-mentioned algorithms and the actual charging load are compared, and the comparison results are shown in Table 1. It can be seen that compared with several other methods, the RMSE and MSE index values ​​of the scene data generated by the algorithm proposed in this article are the smallest, and the R2 value is also the closest to 1, indicating that the method proposed in this article has better evaluation results than other methods, and can simulate the historical ship charging load curve to the greatest extent, proving that the method proposed in this article can effectively learn the historical charging load characteristics. And the calculation time used by the algorithm in this article is also the shortest.

[0110]

[0111] The present invention also provides an embodiment of a computer system, which includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, one or more steps of the above-mentioned method for generating charging load scenarios for electric ships based on DW-CGAN are implemented.

[0112] The present invention also provides an embodiment of a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, one or more steps of the above-mentioned method for generating charging load scenarios for electric ships based on DW-CGAN are implemented.

[0113] In summary, the present invention introduces a non-Markov process to process the forward process of the diffusion model, which can significantly improve the forward noise efficiency of the diffusion model. In the reverse denoising process, the conditional generative adversarial network is used for implicit modeling and reasoning learning, which not only solves the instability problem in GAN training, but also maintains the advantages of the diffusion model in generating high-quality samples and stable training.

Claims

1. A model for generating charging load scenarios for electric ships, characterized in that: include: Diffusion model, using its forward diffusion process to add Gaussian noise to the charging load time series data to reconstruct it to generate noise data that conforms to Gaussian distribution; The conditional generative adversarial network, whose generator uses the denoising function to reversely remove the Gaussian noise in the noise data, thereby restoring the latent probability distribution characteristics of the charging load time series data, and realizing the generation of electric ship charging load scenarios.

2. The electric ship charging load scenario generation model according to claim 1 is characterized in that: The charging load time series data includes charging start time, end time, charging power and charging location / charging pile number.

3. The electric ship charging load scenario generation model according to claim 1 is characterized in that: The Gaussian noise is added by using the non-Markov process noise adding method.

4. A training method for generating a model for charging load scenarios of electric ships, characterized in that: include: Construct a training set, which includes the charging load time series data of each charging pile and the corresponding environmental data; Gaussian noise is added to the training set through the forward diffusion process of the diffusion model to reconstruct it so as to generate noise data that conforms to the Gaussian distribution; Training conditional generative adversarial networks: Noise data and random noise are input into the generator of the conditional generative adversarial network, so that the generator can learn the distribution law of the training set and establish a denoising function that can remove Gaussian noise in the noise data and thus restore the latent probability distribution characteristics of the training set. The training set and the new data generated by the generator using the denoising function are input into the discriminator for distinction. During the training process, the generator and the discriminator continuously optimize their parameters. The training ends when the discriminator determines that the Nash equilibrium is reached between the charged training set and the new data.

5. The training method for generating a model for charging load scenarios of electric ships according to claim 4 is characterized in that: Environmental data includes wind speed, wind direction and temperature.

6. The training method for generating a model for charging load scenarios of electric ships according to claim 4 is characterized in that: During the training process, the Wasserstein distance is used to measure the gap between the distribution of new data generated by the generator and the actual distribution of the training set.

7. A method for generating a charging load scenario for an electric ship, characterized in that: The charging load time series data and random noise are input into a model trained using the training method for generating a charging load scenario model for an electric ship as described in claim 3, and the model generates a charging load scenario.

8. The method for generating a charging load scenario for an electric ship according to claim 7, characterized in that: The charging load time series data is predicted by the historical charging load time series data and the corresponding environmental data.

9. A computer system, characterized in that: include: A processor; a memory, comprising one or more computer program modules; wherein the one or more computer program modules are stored in the memory and configured to be executed by the processor, and the one or more computer program modules include instructions for implementing the training method for the electric ship charging load scenario generation model described in claims 4-6 and the electric ship charging load scenario generation method described in claims 7-8.

10. A computer-readable storage medium for storing non-transitory computer-readable instructions, characterized in that: When the non-temporary computer-readable instructions are executed by a computer, the training method for the electric ship charging load scenario generation model described in claims 4-6 and the electric ship charging load scenario generation method described in claims 7-8 can be implemented.

Citation Information

Patent Citations

  • Economic dispatching optimization method for microgrid comprising electric vehicle charging load

    CN111008748A

  • Method for analyzing voltage stability of power distribution network containing electric automobile

    CN114707292A

  • Discrete probability and CGAN fused power grid random scene generation method

    CN116780509A

  • Image segmentation method and device based on diffusion model and generative adversarial network

    CN117314930A

  • Method for estimating endurance mileage of electric ship

    CN117349959A

Cited By

  • Electric vehicle load gridding prediction method, device, equipment, medium and product

    CN120978715A