PaperlayerView exact revision on arXiv

Source-linked AI summary

Model-Free Renewable Scenario Generation Using Generative Adversarial Networks

Yize Chen, Yishen Wang, Daniel Kirschen, Baosen Zhang

arXiv:1707.09676v2cs.LGeess.SYmath.OC

TL;DR

High renewable penetration requires scenario generation that represents stochastic, spatial, and temporal renewable behavior, while conventional probabilistic approaches can be difficult to scale and sample. The paper uses GANs—two interconnected deep neural networks—to generate scenarios directly from historical data, including condition-specific cases. Results on wind and solar data show realistic, diverse profiles and preservation of spatial and temporal correlations, with efficient feedforward generation.

  • Problem

    High renewable penetration creates a need to model uncertain renewable generation, but model-based approaches face complex dynamics, statistical assumptions, and difficult high-dimensional sampling.

  • Method

    The paper trains interconnected generator and discriminator networks to transform easily sampled noise into renewable scenarios without explicitly modeling the historical distribution.

  • Results

    The method generates realistic and diverse wind and solar profiles while capturing spatial and temporal correlations, including under moderate added noise.

  • Takeaways & Limitations

    Feedforward GANs provide an efficient, scalable, and flexible approach for generating renewable scenarios and condition-specific samples.

Abstract

from arXiv · show

Scenario generation is an important step in the operation and planning of power systems with high renewable penetrations. In this work, we proposed a data-driven approach for scenario generation using generative adversarial networks, which is based on two interconnected deep neural networks. Compared with existing methods based on probabilistic models that are often hard to scale or sample from, our method is data-driven, and captures renewable energy production patterns in both temporal and spatial dimensions for a large number of correlated resources. For validation, we use wind and solar times-series data from NREL integration data sets. We demonstrate that the proposed method is able to generate realistic wind and photovoltaic power profiles with full diversity of behaviors. We also illustrate how to generate scenarios based on different conditions of interest by using labeled data during training. For example, scenarios can be conditioned on weather events~(e.g. high wind day) or time of the year~(e,g. solar generation for a day in July). Because of the feedforward nature of the neural networks, scenarios can be generated extremely efficiently without sophisticated sampling techniques.

I. INTRODUCTION

High renewable penetration makes uncertainty modeling essential for power-system decisions, yet conventional scenario-generation methods struggle with complex spatiotemporal dynamics, statistical assumptions, and high-dimensional sampling. The paper motivates a data-driven GAN approach that generates scenarios directly from historical data without explicitly fitting probability distributions.

  • Motivation: High renewable penetration makes accurate uncertainty modeling important for operation, scheduling, planning, and uncertainty-aware power-system decisions.Scenario sets support stochastic dispatch and unit commitment, renewable-storage operation, and trading strategies.
  • Challenges: Model-based scenario generation is difficult to apply and scale because weather dynamics, nonlinear bounded conversion, and spatial-temporal interactions are complex.These challenges become especially pronounced when multiple renewable power plants are modeled.
  • Challenges: Statistical assumptions in conventional models may be difficult to test, while sampling high-dimensional non-Gaussian distributions is also challenging.The cited literature includes copula, time-series, and machine-learning approaches with varying preprocessing and sampling requirements.
  • Motivation for the proposed approach: The proposed generative approach creates new scenarios directly from historical data without explicitly specifying a model or fitting probability distributions.It also uses unsupervised learning, avoiding manual labeling for large datasets.
  • Observed renewable behavior: Historical and generated wind and solar profiles exhibit rapid variation and strong diurnal patterns characteristic of renewable power generation.The figure compares true historical curves with generated scenarios for both resources.

B. Proposed Method and Main Contributions

The paper adapts GANs for renewable scenario generation using adversarially trained deep neural networks, with generated profiles intended to preserve historical renewable characteristics without memorization. Its contributions include scalable spatiotemporal generation, conditional scenarios, and efficient implementation.

  • Scenario Generation: The method generates distinct wind and solar scenarios that capture rapid variations, strong diurnal cycles, and the intrinsic features of historical data.The examples are intentionally not exact matches to historical trajectories, supporting generation rather than simple memorization.
  • Proposed Method: GANs use a generator and discriminator in a two-player minimax game to learn renewable scenario characteristics from historical data.The generator transforms noise into samples, while the discriminator distinguishes generated samples from historical data.
  • Main Contributions: The paper proposes model-free, data-driven, scalable generation that captures spatial and temporal correlations across renewable power plants.The formulation covers single resources, multiple correlated resources, and event-conditioned generation tasks.
  • Main Contributions: Conditional scenario generation uses labels during training to produce scenarios with specified characteristics, such as high-wind days or seasonal solar output.The procedure is described as adjustable to different conditions of interest.
  • Main Contributions: The paper reports efficient algorithms that can be trained with little or no manual adjustment and scaled to generate large, diverse renewable-profile sets.The code and data are publicly available for the described work.

A. Single Time-Series Scenario Generation

The paper formulates renewable scenario generation as learning the unknown distribution of historical power observations with two jointly trained neural networks. The GAN framework generates realistic scenarios while targeting temporal and spatial correlations and marginal distributions across renewable resources.

  • A. Single Time-Series Scenario Generation: The objective is to generate scenarios that reproduce the stochastic processes and diverse modes represented in historical renewable power data.The formulation uses historical data from N renewable-resource sites as the training set.
  • A. Single Time-Series Scenario Generation: For multiple geographically close sites, generated scenarios should capture temporal and spatial correlations alongside each resource’s marginal distribution.The approach simultaneously generates scenarios for a group of sites sharing the same time horizon.
  • A. Single Time-Series Scenario Generation: The generator transforms sampled noise into realistic scenarios, while the discriminator scores whether inputs resemble real historical data.The two networks are trained simultaneously: the generator upsamples noise, and the discriminator distinguishes real samples from generated samples.
  • A. Single Time-Series Scenario Generation: Training forms a two-player minimax game in which the generator and discriminator optimize losses based on discriminator scores for generated and real samples.The generator seeks higher realism scores for generated samples, while the discriminator separates generated data from historical data.
  • A. Single Time-Series Scenario Generation: The Wasserstein formulation measures distributional discrepancy through the cost of transporting generated samples toward the historical-data distribution.The paper uses empirical means to estimate the expectations and interprets convergence as minimizing this transformation cost.

B. Conditional GANs

Conditional GANs extend unconditional scenario generation by supplying labels that represent event properties of training samples. The generator is then trained to produce scenarios conforming to the selected class of conditions.

  • B. Conditional GANs: Unconditioned generation does not control which specific types of renewable scenarios are produced.The paper motivates conditioning with examples including calm intermittent-wind days and windy days with farms at full load.
  • B. Conditional GANs: Conditional GANs combine event labels with training samples so the generator produces scenarios under a requested class.The label y encodes different classes of conditions.
  • B. Conditional GANs: Labels can be assigned using user-defined metrics such as mean daily power generation or the month from which training samples originate.The paper applies labels based on mean values and seasonal information.
  • B. Conditional GANs: During training, conditional samples with the same label are intended to exhibit similar properties in generated scenarios.The conditional objective uses historical data represented as {x_j|y_j}.
  • B. Conditional GANs: The conditional setup modifies the unconditional GAN model while retaining differentiable neural networks trained with standard gradient-based methods.The networks use multilayer perceptrons, convolutions, normalization, max-pooling, and ReLU units; training uses batch updates and RMSProp.

IV. RESULTS

The results section evaluates wind and solar scenario generation and reports that generated scenarios are visually indistinguishable from historical samples and share their statistical properties.

  • IV. RESULTS: The evaluation examines wind and solar scenario-generation setups, first testing visual similarity and then comparing statistical properties.The section presents the algorithm across several different setups.
  • IV. RESULTS: The reported training procedure uses conditional GAN parameters including learning rate, clipping parameter, batch size, and discriminator updates per generator iteration.Algorithm 1 initializes discriminator and generator weights before alternating their updates.
  • IV. RESULTS: Each discriminator update samples a batch of historical data and a batch of Gaussian noise, then applies gradient descent, RMSProp, and weight clipping.The generator is updated after the discriminator steps using its corresponding gradient and RMSProp update.
  • IV. RESULTS: Generated scenarios are visually indistinguishable from real historical samples and exhibit the same statistical properties.The authors interpret these findings as evidence for an efficient, scalable, and flexible scenario-generation approach.

A. Data Description

The study uses NREL wind and solar integration data to train and test scenario-generation models for multiple renewable sites in Washington.

  • A. Data Description: The dataset contains five-minute power-generation data for 24 wind farms and 32 solar power plants in Washington State.The study constructs training and validation datasets from the NREL Wind and Solar Integration Datasets.
  • A. Data Description: Daily samples are shuffled, with 80% used for training and the remaining 20% used for testing.The split is applied to the selected wind and solar power-generation observations.

B. Model Architecture and Details of Training

The model uses interconnected deep neural networks inspired by DCGAN and Wasserstein GAN architectures, with convolutional and deconvolutional layers for scenario discrimination and generation. Training uses alternating discriminator-generator updates and produces realistic power-generation samples after convergence.

  • Model architecture: The architecture is inspired by DCGAN and Wasserstein GAN, using convolutional layers in the discriminator and deconvolutional layers in the generator.The generator upsamples noise, while the discriminator downsamples scenarios and produces a sigmoid output constrained to [0,1].
  • Training details: The GANs are trained with RMSProp, mini-batches of 32, batch normalization, and alternating four discriminator updates for each generator update.Weights are initialized from a centered Normal distribution with standard deviation 0.02, and batch normalization stabilizes learning.
  • Training outcome: After discriminator convergence, the generator produces realistic power-generation samples, with D(G(z)) and D(x) reaching similar output values.The training procedure alternates optimization of the discriminator and generator until their outputs become comparable.
  • Comparison method: A Gaussian copula is implemented as a comparison method by estimating interdependence from historical observations and sampling from a multivariate normal distribution.The copula approach recursively estimates a covariance matrix across sites before generating scenarios.

C. Scenario Generation

The GAN generates renewable scenarios that resemble held-out wind and solar data while preserving diverse temporal and statistical behaviors. Additional tests examine marginal distributions, spectral structure, and robustness to contaminated training data.

  • Training validation: After 15,000 iterations, the Wasserstein distance between real and generated samples converges near 0, and training continues to 35,000 iterations to demonstrate stability.Early training allows the discriminator to distinguish generated from historical samples, whereas later training reduces their Wasserstein distance.
  • Scenario diversity: Generated wind and solar scenarios closely resemble held-out validation samples while exhibiting diverse modes such as peaks, diurnal variations, and fast power ramps.The validation samples were not used during training, and the generated profiles are intended to capture intrinsic features without memorizing them.
  • Statistical resemblance: The generated scenarios reproduce historical temporal and marginal statistical behavior, with similar autocorrelation coefficients and nearly overlapping CDFs for wind and solar.The autocorrelation comparison evaluates temporal dependence, while the CDF comparison evaluates marginal distributions.
  • Spectral behavior: Generated wind and solar scenarios closely follow the overall shape of historical power spectral density observations across periods from 6 days to 2 hours.PSD is used to evaluate spectral energy distribution, periodic components, and temporal correlation.
  • Robustness to bad data: With 5% wind samples contaminating the solar training set, midnight solar generation remains below 0.3MW relative to a simulated maximum solar output of 16MW.Increasing wind contamination causes more midnight generation, but the reported 5% case remains below 0.3MW and is described as robust to out-of-distribution data.

D. Spatial Correlation

The multi-site GAN represents renewable scenarios as data matrices so that generated wind-farm profiles preserve spatial and temporal correlations. Spatial-correlation tests show agreement with training data under clean and moderately noisy conditions, while very low signal-to-noise ratios degrade quality.

  • Multi-site representation: The multi-site setup feeds the GAN an N × T matrix, using N = 24 wind farms and T = 24 hourly timesteps per scenario.This matrix representation differs from using a single-site diurnal profile and supports joint spatial-temporal scenario generation.
  • Spatial-temporal preservation: Generated one-day wind-farm outputs retain both spatial and temporal correlations visible in historical data not used during training.The comparison is made by visual inspection of groups of real and generated scenarios.
  • Correlation validation: Generated spatial-correlation coefficients agree with training sets even under complex spatial-correlation patterns.The evaluation visualizes correlation-coefficient matrices for clean and noise-contaminated training conditions.
  • Noise robustness: Moderate noise preserves spatial-correlation learning, whereas very low signal-to-noise ratios lead to poor generated-scenario quality.Gaussian noise with standard deviations of 0.01 and 0.1 corresponds to approximate noise-to-signal ratios of 3.38% and 33.8%, respectively.

E. Conditional Scenario Generation

Conditional labels allow the GAN to generate wind scenarios matching distributions associated with specified mean values, ramp events, and forecast-error conditions. The generated samples are realistic and can represent targeted high-wind behavior better than Gaussian copula scenarios.

  • Conditional scenario generation: The GAN conditions wind-scenario generation on mean value, 5-minute ramp events, and forecast-error level using labeled training data.The labels are incorporated during training so generated samples correspond to selected conditions.
  • Conditional scenario generation: Conditional GAN samples are realistic across the practical applications examined, and their overall marginal distributions are compared with validation data.Figure 10 compares conditional generated distributions with validation data and Gaussian copula scenarios.
  • Wind power mean values: 2,000 wind scenarios are generated with 400 samples in each month for the mean-value conditioning experiment.Wind samples are classified into five groups according to their mean generation value.
  • Wind power mean values: Generated samples follow the marginal distributions of validation samples within each wind mean-value class.For µ(x) ≥6, over 30% of generated values lie in [14.4MW,16MW], supporting targeted high-wind-day simulation.
  • Wind power mean values: Gaussian copula scenarios fail to represent the marginal distribution for larger wind-power mean values, where realizations exhibit greater variability.The passage attributes this difference partly to larger fluctuations at higher sample means.

2) Wind Power Ramp Events:

The ramp-event experiment conditions wind scenarios on discretized ramp magnitudes and forecast-error classes. Generated scenarios preserve class-specific distributions, while larger forecast errors are associated with larger generated and observed power values.

  • Wind Power Ramp Events: Wind ramp events are classified into four magnitude classes from below 4.0MW to 12.0–16.0MW for conditional scenario generation.The classes assign labels to training scenarios before GAN training.
  • Wind Power Ramp Events: For intense ramps in classes 3 or 4, generated wind samples have smoother distributions and more power than samples from classes 1 or 2.Within each ramp class, generated scenarios nearly match the marginal distribution of validation forecast samples.
  • Wind Power Forecast Errors: Larger forecast errors coincide with larger power-generation values in both generated scenarios and original forecasts.The passage links these larger errors to greater difficulty in accurately forecasting wind generation at higher output.
  • Wind Power Forecast Errors: Gaussian copula forecast-error scenarios do not capture the overall distribution as well as the proposed conditional GAN approach.The comparison concerns the distribution of forecast-error scenarios under the examined conditions.
  • Comparison with Gaussian copula: Gaussian copula sampling produces more scattered power values because it explicitly estimates a copula and uses a random-number generator.The resulting distributions are less concentrated in the intervals represented by validation data across three simulation groups.

4) Seasonal Solar Power Generation:

Month labels enable the GAN to generate solar scenarios reflecting seasonal variation in daily power and generation duration. The generated samples reproduce the seasonal patterns identified in the validation data.

  • Seasonal Solar Power Generation: Generated solar samples correctly identify Washington’s dry-summer Mediterranean seasonal pattern in both power and duration.The generated data show significant monthly differences in daily power and agree with seasonal sunshine-duration variation.
  • Method and scope: The proposed GAN directly generates renewable scenarios conforming to historical-data distributions without explicitly modeling those distributions.The approach is data-driven and model-free, using deep neural networks and historical data.
  • Method and scope: Retraining with historical data from multiple sites allows GANs to generate scenarios with corrected spatiotemporal correlations without additional tuning.The case study reports successful scenario generation for both wind and solar.
  • Method and scope: A feedforward neural-network structure avoids sampling complex high-dimensional processes and can scale to many uncertainties.The authors state that the methodology requires no particular statistical assumptions.
Loading 1707.09676v2…