Grid Frequency Prediction Using LSTM for Wind Park Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for predicting utility grid frequency are not always accurate and reliable, particularly in the context of wind parks, where fluctuations in wind power generation can destabilize the grid frequency.
Innovation Solution
A method using a recurrent neural network (RNN), specifically a Long Short-Term Memory (LSTM) network, to predict grid frequency by combining utility grid and wind park measurement data, including active and reactive power, to improve accuracy and reliability of frequency predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional statistical methods (polynomial regression, ARMA, ARIMA) are used for frequency prediction, then the prediction method is simple, but the prediction accuracy is insufficient
Solution Approach 1:
The patent transitions from conventional statistical methods to a recurrent neural network (RNN) with long short-term memory (LSTM) layers, fundamentally changing the prediction approach from traditional statistical models to deep learning. This parameter change enables the system to capture complex temporal dependencies in wind power and frequency data, significantly improving prediction accuracy while maintaining computational feasibility through the LSTM architecture's efficient handling of sequential data
2Device complexity
If simple artificial neural network models are used for frequency prediction, then the model structure is simple, but the prediction reliability is insufficient under varying wind conditions
Solution Approach 1:
The patent employs a recurrent neural network with LSTM layers that dynamically adapt to changing wind conditions. The LSTM architecture maintains internal memory states that allow the model to learn and respond to temporal patterns in wind power fluctuations, making the prediction system reliable under varying operational conditions rather than being static like simple neural networks
Solution Approach 2:
The RNN architecture inherently incorporates feedback mechanisms through its recurrent connections, where past predictions and actual measurements inform future predictions. This feedback loop enables the model to continuously improve its predictions by learning from historical data patterns, enhancing reliability under different wind scenarios
3Adaptability or versatility
If high penetration of renewable generation is used to supply electrical power, then the sustainability is improved, but the grid frequency stability is endangered
Solution Approach 1:
The patent applies preliminary action by predicting grid frequency deviations before they occur using the trained RNN model. By analyzing historical wind power and frequency data, the system forecasts future frequency states, allowing transmission system operators to take preventive measures (such as adjusting balancing power or generator output) before actual frequency deviations occur, thus maintaining stability despite high renewable penetration
Solution Approach 2:
The prediction system acts as an intermediary between renewable generation and grid frequency control. It processes wind power data and translates it into frequency predictions, enabling operators to make informed decisions about balancing power and generator control, thereby mediating the impact of variable renewable generation on grid frequency stability
Data Source
AI summary
A method of predicting a frequency value of a utility grid to which a wind park is connected is provided, the method including: obtaining plural utility grid measurement values pertaining to a predetermined time range before and until a present point in time; obtaining plural wind park measurement values pertaining to the time range; feeding the plural utility grid measurement values and the plural wind park measurement values into a recurrent neural network trained to output the frequency value at at least one next point in time, the next point in time being in particular between 0.5 s and 2 s after the present point in time.