Innovative Control Strategy for Wind Storage Systems Utilizing Temporal Pattern Attention and Bidirectional Gated Recurrent Units

“`markdown
Increasing the penetration of wind power significantly affects the operational mechanisms of power systems. To enhance the controllability of wind farms and align it with that of traditional energy units, this paper introduces a control strategy for wind storage systems that utilizes temporal pattern attention (TPA) and bidirectional gated recurrent units (BiGRUs).

The proposed strategy employs BiGRU to analyze the time series data from the energy storage output, the actual output of the wind farm, and the energy storage state, thereby enhancing the stability of the wind storage system’s control. Additionally, TPA is integrated to assign varying weights to the hidden states in the neural network, emphasizing the importance of local time series data on the current energy storage output, effectively improving model performance and reducing control deviation. The efficacy and stability of this control strategy are validated using actual data from a wind farm, showing a significant improvement in the economic performance of the wind storage system.

Keywords

Bidirectional gated recurrent unit; control strategy; data-driven; deep learning; wind storage combined system.

1. Introduction

Joint control strategies for wind storage systems are essential for improving the competitiveness and regulation of wind power in markets with high penetration levels. Energy storage systems can achieve dispatchability comparable to conventional units by adjusting the active output of wind power to meet dispatch schedules. Given the cost and capacity constraints of energy storage, optimizing the control strategy to enhance output accuracy and economic efficiency is crucial.

Much research has focused on the joint control strategy of wind storage systems, primarily categorizing them into two types: direct control strategies and process optimization control strategies.

Direct control involves real-time adjustments based on deviations between the actual and planned outputs of the wind farm. This includes methods like mode decomposition, proportional–integral–derivative (PID) control, and fuzzy control. While direct control can quickly respond to errors, it may lead to issues such as overcharging or over-discharging of the energy storage system during significant fluctuations in wind power, adversely affecting control efficiency and cost.

Process optimization strategies aim to minimize errors by establishing mathematical models for energy storage orders and throughput as objective functions, continually refining the energy storage output. However, these strategies may struggle to converge in complex environments, which can compromise control effectiveness.

Integrating deep learning methods with these two control approaches can leverage their strengths, enhancing stability and overall performance. Deep learning excels at handling the stochastic nature of wind power and has shown promise in various power system applications, such as state estimation and fault detection.

2. Structure and Control Criteria

The joint control strategy structure for wind storage systems is illustrated in Figure 1. The regulation power calculation is based on the actual output of wind turbines, the planned output, and the energy storage state of charge (SOC). A predetermined dead zone defines when the energy storage system should activate, only responding when control deviations exceed these thresholds.

To integrate large-scale wind power into grid operations safely and efficiently, the active power regulation capacity of grid-connected wind power must meet specific management requirements. For instance, the China Southern Power Grid stipulates that the assessment is conducted every 15 minutes, with conventional units allowed a power deviation rate not to exceed ±2.5%.

The evaluation criterion for wind power active output deviation is defined as:
[ K_1 = \frac{1}{n} \sum_{i=1}^{n} \frac{P_{P_i} – P_{A_i}}{C_{cap}} ]
where ( C_{cap} ) is the wind farm’s starting capacity, ( P_{P_i} ) and ( P_{A_i} ) are the planned and actual outputs at sampling point ( i ), and ( n ) is the number of sampling points.

3. Designing Control Strategy and Model

The control process for the wind storage strategy, based on advanced rolling optimization control and PI control, is divided into two parts. The first part determines the type of energy storage action, while the second calculates the specific charge and discharge values based on the first part’s findings.

Three trained TPA-BiGRU models are implemented to facilitate these functions. The classification model identifies energy storage states (charging, discharging, inactive) through one-hot encoding, while the regression models compute the power values for charging and discharging.

The BiGRU network processes wind power and energy storage data to extract temporal relationships, and the TPA mechanism enhances the model’s memory function, emphasizing local information relevance to current energy storage outputs.

4. Training Process of the Model

The proposed model consists of a deep BiGRU neural network and TPA, requiring energy storage action data generated through optimal and PI control for training. The input variables include actual and planned power outputs and past SOC data.

Data preprocessing involves standardizing continuous data and applying one-hot encoding to discrete states. The TPA-BiGRU model training necessitates three-dimensional data for supervised learning, achieved through a sliding window approach.

5. Case Study

Using TensorFlow v2.18.0, the model is tested on real historical data from a 100-MW wind farm over 33 days, with specific parameters derived from existing literature. Five control strategies are compared for effectiveness: TPA-BiGRU, TPA-BiLSTM, BiGRU, AROCS, and traditional PI control.

Results demonstrate that TPA-BiGRU not only maintains the lowest RMSE and energy storage actions but also extends battery life and operational efficiency, further enhancing competitiveness in the energy market.

6. Conclusions

This study presents a joint control strategy for wind storage systems grounded in TPA-BiGRU, which improves both economic and operational stability. The model can adapt to the complex operational conditions of wind farms while delivering real-time control solutions.

Future research directions include enhancing adaptive hybrid control strategies, incorporating transformer architectures for better context awareness, ensuring resilient operations under extreme disturbances, and optimizing scalability for larger systems.
“`