Models and Methods of Wind Speed Forecasting Essay Example

The reduction of energy-related carbon dioxide emissions is central to the energy revolution. If the world is to meet the agreed-upon climate targets, it must rapidly transition away from the consumption of fossil fuels that cause climate change and toward cleaner, renewable forms of energy. Wind energy is an integral component of this critical environmental and energy solution. [1]. Its unpredictable and intermittent qualities complicate power system operation, particularly unit commitment and day-ahead power generation scheduling. Forecasting renewable energy output with high accuracy may significantly minimize the risk of power system failure [2-3].

Wind forecasting models can be classified as very short-term (few seconds to 30 minutes ahead), short-term (30 minutes to 6 hours ahead), medium-term (6 hours to 1 day ahead), and long-term (1 day to 1 week or more ahead) models. Ultra-short-term forecasts aid in real-time turbine control and load tracking; short-term forecasts aid in load dispatch planning; medium-term forecasts aid in energy trading and power system management; and long-term forecasts advise optimal maintenance scheduling [4].

In general, the current literature has proposed three types of wind speed forecasting models: 

• Some employed physical approaches for wind forecasting, such as numerical weather prediction (NWP) [6–7]. NWP forecasts weather using hydro- and thermodynamic models of the atmosphere, considering specified beginning values and boundary conditions. Wind forecasting becomes erroneous when the precision of the NWP is poor. 
• Statistical approaches were also utilized for wind energy forecasting. There is probabilistic forecasting [8], polynomial autoregressive model [9], and Sparse Bayesian learning [10] for this kind of approach. These techniques focus on the relationship between wind forecasts and explanatory variables, but they lack adaptive and learning capabilities.
• Artificial Intelligence (AI) methods employ Machine Learning and Deep Learning techniques. Some of these techniques are XGBoost Tree-Based algorithm [11], Deep Boltzmann Machine (DBM) [12], Autoregressive integrated moving average with Artificial Neural Network (ARIMA-ANN) [13], General Regression Neural Network (GRNN) [14], Multi-scale Long-short Term Memory (LSTM-M) [15], and Convolution Neural Network [16]. These approaches have the benefit of not requiring a preset mathematical model. However, these methods are only effective for extremely short-term forecasting and necessitate the use of additional feature extraction techniques.

According to the above discussions, artificial intelligence is the most popular to conduct wind speed forecasting; however, they have some limitations:

• Most of the methods used [11]-[15] only use one location to produce a 24-hr ahead wind speed forecasting. The data is very limited to a certain location specially when dealing with designing a wind farm.
• Machine learning [11] is prone to overfitting which can provide inaccurate prediction specially for longer hours.
• Several methodologies require additional method like ARIMA and resampling methods to extract features from existing limited meteorological information [13,14]. Although there is a feature extraction, the interpretation can only be done through the given data rather than within the timesteps.
• For a larger prediction horizon, interval prediction was often utilized [11]-[15]. Although the wind speed range is supplied, such a solution cannot be used for further dispatch and scheduling applications.

Although existing techniques are provided, the current trend indicates the use of the concept of Quantum mechanics in Quantum computers. The belief that quantum algorithms can solve some problems more efficiently than classical approaches is a motivating reason for research into quantum computing. However, the line between classical and quantum computing is continually changing as new classical and quantum algorithms disturb the environment, always altering the line that divides these two paradigms. This interaction is beneficial because it allows classical and quantum algorithm creators to benefit from one other's inventions, furthering our understanding of the boundaries of computation [19]. 

This research presents a hybrid deep learning quantum-inspired neural network strategy for improving the accuracy of 24-hour ahead wind power forecasting. This new technique employs CNN in conjunction with LSTM and a Quantum-Inspired Neural Network (QINN). This paper makes the following contributions:

1. Historical geographical and temporal data from the 4 selected sites are used to generate the 2D information used in the proposed method, which does not require any extra digital signal processing such as ARIMA or resampling [13][14].
2. A potential location for an offshore wind project is being investigated. The average wind speed multiplied by 24 hours is used to approximate the location of other stations that provide auxiliary wind speed information. To account for temporal shifts or delays, a maximum cross-correlation analysis is utilized.
3. The proposed method uses a CNN structure with Average Pooling sequential to LSTM for prediction. The CNN provides the features from the spatio-temporal data while the LSTM interprets the features extracted from the data along the timesteps. Deep Learning methods are not too prone for overfitting.
4. The Quantum Delta Potential Well, a variation of Quantum PSO, determines the structural parameters and hyperparameters of the CNN, LSTM, and QINN. This kind of PSO restrains the particle premature convergence and fall into local convergence. 
5. To the best of the authors' knowledge, the suggested technique is the first to be constructed as a QINN based on the mathematical model utilized for the fully connected layer to provide the 24-hr ahead forecast.

This paper is organized as follows. Section II discusses an overview of the theoretical backgrounds of the CNN, LSTM, PSO, and quantum computing. Section III provides the proposed method with the QPSO and QINN. Section IV provides results from the proposed method and a comparison with the related literature. In Section V, the study is concluded with some recommendations.