Medium and Long Term Scenario Generation of Multi-Temporal Renewable Energy Based on GAN

With China’s “double carbon” target, the future power system will show A significant percentage of the energy generated comes from renewable sources, specifically wind power, which has obvious randomness, volatility and intermittency, and will have a greater impact on the safety of the power system. How to effectively characterize the randomness and volatility of its power output is a major challenge for power system dispatch. The method we suggest in this paper utilizes generative adversarial networks to produce medium- and long-term scenarios for wind power generation. This approach allows for the creation of multiple sets of wind power generation data that share the same characteristics as the historical data., Our approach takes into account the spatial correlation between wind power generation at wind farms situated in various geographic locations, fully describing the randomness and spatio-temporal correlation of wind power, and providing an important reference for power system planning and scheduling.

the joint probability density distribution corresponding to the set of time-series scenarios by implicitly modelling the spatio-temporal properties of random variables.Instead of artificially fitting to find parameters, a lot of historical data is simply put into the model for training, and the trained model then has the ability to generate the same features to process the scenes, with generative adversarial networks being the more suitable method [3] .
The literature [4] innovatively proposed Generative Adversarial Networks (GAN).Generative models are expected to make use of large amounts of unlabelled training data.There are two main benefits of utilizing these methods.Firstly, they have the ability to produce fresh scenarios by directly utilizing historical data without requiring a specified model or the adaptation of a probability distribution.Secondly, these methods employ unsupervised learning, eliminating the need for laborious manual annotation and enabling the handling of vast datasets.Generative adversarial networks (GANs), as a class of deep learning models, are well placed to enable the enhancement and generation of power data samples.In the literature [5]，A new approach is suggested for creating wind power output scenarios across multiple regions using conditional generative adversarial networks.The method uses a threedimensional convolutional network to design a network structure suitable for multi-region wind power output scenario generation.Through training the conditional generative adversarial network using a game-based approach, the method acquires an understanding of the patterns in actual wind power output data across multiple regions, as well as the correlation between input and output data.In the literature [6], Our suggested approach employs a conditional adversarial network to produce day-ahead scenarios of renewable energy, incorporating a specialized network architecture and utilizing the Wasserstein distance as the discriminator's loss function.An improved PV power scenario generation method with gradient penalty for generative adversarial networks is proposed in the literature [7].In the initial stage of this method, the Wasserstein distance is employed as the loss function to facilitate the creation of two deep neural networks, the generator and the discriminator, for adversarial training, and secondly introduces a gradient penalty strategy in the loss function to enhance the Lipschitz continuity constraint of the model and innovatively applies it to PV power scene generation, which improves the convergence speed of the scene generation model and the quality of the generated scenes.In other article, Our proposed method involves utilizing an improved conditional deep convolutional generative adversarial network to generate scenarios for power generation based on environmental conditions.The first step involves designing a network structure for the conditional generative adversarial network that is tailored for generating wind power and photovoltaic output scenes, and employing the Wasserstein distance as the discriminator's loss function.Then, the generator is trained to learn the mapping relationship between random noise and the training set of real historical data through game training of the conditional generation adversarial network, so as to efficiently generate scenes close to the distribution of real scenes.The various GAN models described above can describe the prediction error interval under the corresponding prediction value well, but there are also problems such as unstable training, high computational resource consumption and uncontrollable generation results, and they do not only consider temporal or spatial correlation, but not the two in a unified manner.

Advantages of GAN
Compared to other common generative models, GAN has the following advantages: 1) In terms of working mechanism, the generator of GAN can directly sample the data and learn a function that approximates the real data distribution; while the discriminator can directly fit the probability distribution of the sampled data and optimize the parameters of the generator's function, which means that GAN has the full ability to mine the probability distribution of the data samples without the need to know the explicit distribution of the real data samples, and also without more mathematical assumptions.
2) The proposed method employs deep neural networks, such as autoencoder neural networks and convolutional neural networks, as the generator and discriminator in the GAN network structure.This enables the construction of deep generative models that produce higher quality data samples at a faster generation speed compared to non-deep generative models.
3) Regarding the objective function, The objective function of the generator is to minimize the discrepancy between the probability distribution of the generated data and that of the real data.On the other hand, the discriminator's objective function is utilized to evaluate the discrepancy in probability distribution between the generated and real data; GAN enables the model to spontaneously learn the inherent laws of the data samples through its own adversarial training method.

Wasserstein GAN
It has been shown that the original GAN model suffers from training non-convergence.In this process, both the generator and discriminator experience a lack of significant convergence in their loss functions, and there is no direct representation of the training process.
In order to address these issues, a WGAN model based on the Wasserstein distance is used for the generation of integrated energy load scenarios.The Wasserstein distance measures the distance between the actual distribution of the load and the generated distribution during the model's training, and gradually converges as the network parameters are trained.
The Wasserstein distance is mathematically defined as follows: Where: π (x,x) is the joint probability density function satisfying the edge distribution of ( ) p x and ( ) p x′ , and ( , ) d x x is the inter-scene distance measure.Although the WGAN objective function makes it easier to optimise the generator, it may still produce low-quality samples for some specific inputs or fail to converge in some settings.The problems with the WGAN algorithm in terms of convergence speed and convergence stability are improved by introducing a gradient penalty term p G to limit the gradient descent of the generator and discriminator.
Thus, the objective function V(G, D) is converted to

Overview of the algorithm
The scenario generation methods mentioned in this paper are applicable to both wind power and PV, and an example is made for both wind power and PV.Validation analysis was conducted using real output data obtained from a wind farm and PV plant located in the East China Power Grid.The granularity of the data is 15 minutes, and there are 366 sets of data, each set of data is the wind or PV output data at 96 hours a day.The algorithm's generative adversarial network model was constructed using the deep learning framework TensorFlow, and the model was trained using the CPU.The algorithm has been configured with a learning rate of 0.00001 and 5000 training iterations.

Model training
The loss function is based on the Wasserstein distance metric.A normally distributed noise of length 100 is generated as input, and the generator continuously extracts features between the training set data to continuously strengthen its own generative skills, and the discriminator also continuously improves its own discriminative skills, and finally the generator is trained.The parameters are in table1,table2

Generating scenario analysis
From the above figure, the results demonstrate that the generative adversarial network is capable of effectively generating wind and PV output scenarios, as the generated images exhibit similar characteristics to the original output maps.On this basis, the annual wind power or PV installation planning values for a certain year in the future can be multiplied to generate medium and long-term scenery output scenarios.The specific results are shown in the Figure 2,Figure 3,Figure 4,Figure 5 .

Conclusion
The uncertainty in power system generation has become more complex and multifaceted.Accurate and efficient generation of new energy generation scenarios in their application to power system optimisation decision problems, which are of great importance for system safety, economic operation and planning.
The present study introduces a novel approach for generating medium to long-term scenarios of new energy generation, specifically designed for the planning and scheduling of power systems.The method NESP-2023 Journal of Physics: Conference Series 2592 (2023) 012027 IOP Publishing doi:10.1088/1742-6596/2592/1/0120278 uses generative adversarial networks to extract the spatio-temporal correlation between data and generate scenic generation scenarios with such spatio-temporal correlation.The example considers two representative new energy generation methods, wind power and photovoltaic power, and compares the actual scenery generation scenarios with the scenery generation scenarios generated by GCN to obtain the expected results.The findings of this study, highlighting its potential for modeling and predicting new energy power systems.These results hold significant implications for advancing the field of new energy power system research.

Figure 1 .
Figure 1.Basic structure of GAN The generated sample G(z) and the real sample x subject to the probability distribution ( ) data P x are used as inputs to the discriminator D. The output of the discriminator D is a scalar D(G(z)) representing the probability of the generated sample G(z) subject to the real distribution ( ) data P x The training goal of the generator is to generate sample G(z), ( ) G P Z .During training, the objective of the discriminator is to correctly distinguish between real and generated samples.The output of the discriminator is then utilized to optimize the network structures of both the generator and discriminator through the gradient function.Following adversarial training, the generator and discriminator reach a Nash equilibrium state where the discriminator D is unable to discern between real and generated samples.At this point, it is assumed that the generator G has successfully learned the probability distribution of the real samples.both the generated sample G(z) and the real sample of the generator G obey the probability distribution ( ) data P x .Considering the training goals of the generator G and the discriminator D, the loss functions of the generator and discriminator G L and D L are shown in equations (1.1) and (1.2) respectively.G

Figure 2 .
Figure 2. Original wind power output scenarios

Figure 4 .
Figure 4. Original photovoltaic power output scenarios