Deep Learning for Modeling and Dispatching Hybrid Wind Farm Power Generation

TL;DR

The paper tackles maximizing wind-farm value by combining a per-farm data-driven dispatch policy with a probabilistic wind-power generation model, both built on LSTM-based architectures. It introduces COVE-NN, an unsupervised, storage-aware dispatch framework, and an NQF-RNN based wind power model, validated on the Pyron and Palouse wind farms with substantial gains over baselines. Key results include a ~32% improvement in annual COVE for Pyron and a 9.5% RMSE reduction with better power-curve similarity for Palouse, demonstrating the practicality of site-specific, data-driven control and generation modeling. The work lays a foundation for integrating generation modeling with dispatch decisions, potentially extending to other renewables and grid-scale applications.

Abstract

Wind farms with integrated energy storage, or hybrid wind farms, are able to store energy and dispatch it to the grid following an operational strategy. For individual wind farms with integrated energy storage capacity, data-driven dispatch strategies using localized grid demand and market conditions as input parameters stand to maximize wind energy value. Synthetic power generation data modeled on atmospheric conditions provide another avenue for improving the robustness of data-driven dispatch strategies. To these ends, the present work develops two deep learning frameworks: COVE-NN, an LSTM-based dispatch strategy tailored to individual wind farms, which reduced annual COVE by 32.3% over 43 years of simulated operations in a case study at the Pyron site; and a power generation modeling framework that reduced RMSE by 9.5% and improved power curve similarity by 18.9% when validated on the Palouse wind farm. Together, these models pave the way for more robust, data-driven dispatch strategies and potential extensions to other renewable energy systems.

Figures (8)

• Figure 1: Baseline Network for Adaptation to Problem Constraints
• Figure 2: NQF-RNN Model Architecture. The modification to the common architecture is highlighted in blue, whereby an additional quantile level input $\alpha$ is passed to the feedforward network, known as the neural quantile function (NQF).
• Figure 3: COVE-NN Model Architecture. The modification from the common architecture is highlighted in blue, whereby outputs from the network are passed through Algorithm \ref{['alg:post']}. The resulting $r_t'$ values are recorded and the $s_{t+1}$ values are forwarded recurrently to the LSTM alongside the $t+1^{th}$ set of input covariates.
• Figure 4: Time series of historical data vs. model predictions
• Figure 5: Log-transformed power curve densities. (a) Historical data. (b) Model predictions. (c) Density difference between historical data and model predictions