Assessment of Offshore Wind Power Potential and Wind Energy Prediction Using Recurrent Neural Networks

Abstract

In recent years, Taiwan has actively pursued the development of renewable energy, with offshore wind power assessments indicating that 80% of the world’s best wind fields are located in the western seas of Taiwan. The aim of this study is to maximize offshore wind power generation and develop a method for predicting offshore wind power, thereby exploring the potential of offshore wind power in Taiwan. The research employs machine learning techniques to establish a wind speed prediction model and formulates a real-time wind power potential assessment method. The study utilizes long short-term memory networks (LSTM), gated recurrent units, and stacked recurrent neural networks with LSTM units as the architecture for the wind speed prediction model. Furthermore, the prediction models are categorized into annual and seasonal patterns based on the seasonal characteristics of the wind. The research evaluates the optimal model by analyzing the results of the two patterns to predict the power generation conditions for 1 to 12 h. The study region includes offshore areas near Hsinchu and Kaohsiung in Taiwan. The novelty of the study lies in the systematic analysis using multiple sets of wind turbines, covering aspects such as wind power potential assessment, wind speed prediction, and fixed and floating wind turbine considerations. The research comprehensively considers the impact of different offshore locations, turbine hub heights, rotor-swept areas, and wind field energy on power generation. Ultimately, based on the research findings, it is recommended to choose the SG 8.0-167 DD wind turbine system for the Hsinchu offshore area and the SG 6.0-154 wind turbine system for the Kaohsiung offshore area, serving as reference cases for wind turbine selection.

1. Introduction

In recent years, Taiwan has actively pursued the development of renewable energy. Taiwan’s diverse renewable energy sources encompass solar photovoltaic, onshore wind power, offshore wind power, geothermal energy, biomass, hydropower, and fuel cells. Despite Taiwan’s limited land area, the availability of flat land for renewable energy development is constrained. As the onshore renewable energy sector approaches saturation, offshore wind power, located at sea, has become a primary focus for energy development in Taiwan. Internationally evaluated to possess 80% of the world’s premier wind sites in the western waters of Taiwan [1], offshore wind power is considered a pivotal energy source in the region.

In 2017, Taiwan’s Energy Administration (EA) launched a 4-year plan for wind power generation. A key objective of this initiative is to attain a target installed capacity of 3 GW in offshore wind power by 2025, accompanied by an annual generation goal of 11 billion kilowatt-hours [2]. With numerous potentially suitable locations identified in the Taiwan Strait for the establishment of offshore wind farms, the assessment of offshore wind power potential and real-time wind energy prediction emerges as a critical undertaking.

Wind farm characteristics are commonly modeled using the Weibull distribution, which is applied to fit wind speed data for the estimation of wind power potential [3,4,5]. Various studies, including those by Altunkaynak et al. [6], Fyrippis et al. [7], and Zárate-Miñano et al. [8], have demonstrated that combining the Weibull distribution with suitable models can effectively estimate wind power potential at different locations. In the specific context of wind turbine siting coupled with the Weibull distribution, Beaucage et al. [9] employed synthetic aperture radar and the Weibull probability density function to calculate offshore wind speed potential in the Saint Lawrence River region of Canada. Subsequently, they conducted wind turbine siting based on wind energy density. This approach showcases the versatility of the Weibull distribution in assessing wind characteristics and informing decisions related to wind energy infrastructure placement.

In the analysis of offshore wind fields, Oh et al. [10] conducted a comprehensive examination of the effective offshore wind field in South Korea utilizing offshore buoy data. They integrated ground meteorological station data with remote sensing QuikSCAT data to generate offshore wind speed density and wind energy density maps. Taking a seasonal modeling approach, Ganea et al. [11] assessed offshore wind energy for the Greek Peninsula by employing meteorological forecast models and satellite remote sensing data. Their study involved the selection of wind turbines to simulate various wind speed scenarios, including annual average, daily average, nighttime average, four-season daily average, and four-season nighttime average wind speeds and wind energy for five regions. González et al. [12] estimated future changes in wind properties in the Canary Archipelago, both in the middle and at the end of this century. The pseudo-global warming method was employed to dynamically regionalize the wind climatology for the Canary Islands, with the Weather Research and Forecasting model chosen as the regional climate model. In the case of Taiwan, Chang et al. [13] utilized long-term data from ground stations to evaluate wind speeds using the Weibull wind speed distribution. They determined wind turbine capacity factors and performance factors for four specific locations. Meanwhile, Cheng et al. [14] established a wind speed statistical model for Central Taiwan, aiming to improve the accuracy of very short-term wind speed predictions. Based on the literature mentioned above, it is recognized that a careful selection of specific wind turbines is crucial for accurately estimating effective wind power generation capacity. In making this selection, the study took into account several factors, including the widespread global usage of certain wind turbine models, the adoption of specific models by offshore wind farms in Taiwan, and the incorporation of active offshore wind turbine demonstration models from other countries. Additionally, the use of buoy data and ground station data can effectively simulate future wind speeds, while the use of seasonal modeling can provide a more accurate analysis of seasonal wind speeds. The calculation of wind speed probability density functions helps derive characteristics of the wind field.

Regarding wind power potential and wind speed prediction, wind speed is considered a stochastic variable in meteorological data. In wind energy prediction, there are primarily three methods: physical, statistical, and soft computing methodologies. Physical models typically utilize numerous physical variables, including geographical and meteorological variables, and are designed for long-term and trend forecasting [15,16,17]. However, due to the substantial input data requirements, the computational time for physical models is often lengthy, and their accuracy depends on numerical weather prediction (NWP). In statistical methods, models use historical data as input to calculate future wind energy data. Commonly employed techniques include autoregressive moving average models and autoregressive integrated moving average models [18,19].

Soft computing methodologies, widely used in renewable energy prediction [20], include artificial intelligence methods, such as artificial neural networks, which are utilized for wind energy prediction [21]. Additionally, the support vector machine has been highlighted by Yang et al. [22] as capable of providing good wind energy prediction results when training samples are limited. Liu et al. [23] adopted the adaptive neural fuzzy inference system, a nonlinear method for wind speed prediction, though it faces challenges in training with large datasets. In recent years, various deep learning algorithms have been developed. For example, a deep belief network can be applied to wind speed and wind energy prediction models with large datasets [24]. The convolutional neural network uses a two-dimensional approach for wind power prediction [25]. Recurrent neural networks (RNN) are models capable of computing long-time series data and have been used in wind speed prediction [26]. Long short-term memory network (LSTM), a branch of RNN, addresses the vanishing gradient or exploding gradient issues in RNN by controlling the calculation of input weights and assisting in updating weights while remembering the necessary training information [27]. For instance, Liu et al. [28] proposed a hybrid model for wind power prediction based on complementary ensemble empirical mode decomposition with adaptive noise, sample entropy, bidirectional long short-term memory network, and Markov chain.

Furthermore, the gated recurrent unit (GRU), which shares structural similarities with LSTM, has been indicated in several studies to outperform LSTM in wind speed prediction [29,30,31]. Moreover, related works, such as [32,33], have employed a stacked RNN with LSTM units. In summary, soft computing methods, particularly machine learning methods, are commonly employed in wind speed prediction. Given the inclusion of long-time series data in this study, LSTM and GRU were chosen as modeling algorithms. Subsequently, a stacked LSTM, consisting of different LSTM layers, is compared with individual LSTM and GRU models.

The main objective of this study is to establish a robust wind speed prediction model and thoroughly assess the wind power potential in the Hsinchu and Kaohsiung regions of Taiwan. These regions hold pivotal economic roles in fulfilling the regional electricity demand, underscoring the indispensability of offshore wind energy for sustainable and renewable power sources. The integration of machine learning techniques, specifically RNN-based models, is aimed at elevating the precision of wind speed predictions, thereby contributing to a comprehensive understanding of offshore wind power dynamics. The secondary objectives are defined as follows:

• Evaluate offshore wind power potential: Conduct a thorough assessment of the offshore wind power potential in the research areas to effectively address future electricity demand. This evaluation places significant emphasis on the role of offshore wind as a sustainable and reliable energy source, crucial for meeting the power requirements of regional economic development.
• Develop real-time wind speed models: Utilize RNN-based models to construct a real-time wind speed prediction model. This model will be formulated using offshore buoy data and ground station meteorological factors as inputs to the machine learning algorithm. The primary goal is to establish accurate and reliable annual and seasonal wind speed prediction models.

2. Research Area and Data

The study focuses on the offshore areas near Hsinchu and Kaohsiung in Taiwan. Figure 1 illustrates the locations of ground stations and buoys in the research area, including Hsinchu Station, Hsinchu Buoy, Kaohsiung Station, and Liuqiu Buoy. Data collection for this study spans from 2005 to 2022, with a data interval of one record per hour.

Table 1. Meteorological attribute statistics for ground stations.

Attribute	Unit	Min–Max, Mean
		Hsinchu Station	Kaohsiung Station
Surface pressure	hPa	964.6–1033.3, 1099.46	976.9–1030.9, 1012.28
Sea level pressure	hPa	967.8–1037.0, 1012.80	977.2–1031.3, 1012.67
Avg wind speed	m/s	0–12.7, 1.96	0–16.6, 2.10
Avg wind direction	°	0–360, 114.05	0–360, 205.75
Max avg wind speed	m/s	0–15.1, 2.67	0–17.2, 2.76
Max avg wind direction	°	0–360, 121.87	0–360, 217.89
Peak gust speed	m/s	0–36.7, 6.12	0–39.1, 4.89
Peak gust direction	°	0–360, 139.61	0–360, 212.59
Precipitation	mm	0–375, 0.23	0–108, 0.22
Duration of rainfall	h	0–1, 0.97	0–1, 0.04

and

Table 2. Meteorological attribute statistics for buoy stations.

Attribute	Unit	Min–Max, Mean
		Hsinchu Buoy	Liuqiu Buoy
Avg wind speed	m/s	0.6–28.6, 6.76	0.4–30.7, 3.94
Avg wind direction	°	0–360, 121.75	0–360, 214.87
Gust wind speed	m/s	1.1–38.2, 8.51	1.1–40, 5.23

present the statistical data for ground meteorological stations and offshore buoy data, respectively. The tables display the maximum, minimum, mean, and standard deviation statistics for various meteorological parameters. The ground station data include surface pressure (PS01), sea level pressure (PS02), average wind speed (WD01), average wind direction (WD02), maximum average wind speed (WD03), maximum average wind direction (WD04), peak gust speed (WD05), peak gust direction (WD06), precipitation (PP01), and duration of rainfall (PP02). The buoy station data include average wind speed (Vm), average wind direction (Dm), and gust wind speed (Vg). The ground station data were measured at a height of 10 m above sea level, while the buoy station data were measured at a height of 3 m above ground level.

3. Wind Power Potential Assessment

The research begins with the assessment of wind power potential, followed by the establishment of a wind speed prediction model. Finally, suitable wind turbines are selected to simulate real-time operation and evaluate turbine performance. The subsequent research process is outlined as follows:

• For the wind power potential assessment, the study utilizes wind turbine blade theory and the Weibull probability density function to estimate wind power potential. Several sets of wind turbines are selected, and the ideal annual power generation is analyzed.
• Next, the wind speed prediction model is established using machine learning algorithms, specifically LSTM, GRU, and stacked LSTM models. Separate models are built for the nearshore areas of Hsinchu and Kaohsiung, creating annual and seasonal prediction models to determine the optimal wind speed prediction model.
• Lastly, the evaluation of real-time turbine performance involves assessing the performance of different turbine models under real-time operation.

3.1. Estimation of Weibull Wind Speed Density

Many studies utilize the Weibull distribution to fit wind speed data for estimating wind energy potential [34,35,36]. The probability density function and cumulative distribution function of the Weibull distribution are expressed as follows:

$$f\left(v\right)=\frac{k}{v}{(\frac{v}{c})}^k\exp [-(\frac{v}{c}{)}^k]$$

(1)

$$F\left(v\right)=1-\exp [-(\frac{v}{c}{)}^k]$$

(2)

where f(v) is the probability density function of wind speed, F(v) is the cumulative distribution function of wind speed, k is the shape parameter (shape factor), c is the scale parameter (scale factor), and v is the wind speed (m/s).

The parameters k and c can be obtained using the method of maximum likelihood. The formulas are as follows:

$$k={\left(\frac{{{\displaystyle \sum }}_{i=1}^n{v}_i^k\ln \left(v_i\right)}{{{\displaystyle \sum }}_{i=1}^n{v}_i^k}-\frac{{{\displaystyle \sum }}_{i=1}^n\ln \left(v_i\right)}{n}\right)}^{-1}$$

(3)

$$c={\left(\frac{1}{n}{\displaystyle {{\displaystyle \sum }}_{i=1}^n}{v}_i^k\right)}^{\frac{1}{k}}$$

(4)

where v_i represents the wind speed at time i, and n is the number of data points for wind speed.

This study utilizes the Weibull distribution to describe the wind speed density functions for two locations: Hsinchu Buoy and Liuqiu Buoy. The estimated parameters (k, c) for Hsinchu Buoy are (2, 9.478), and for Liuqiu Buoy, they are (2, 2.708). Figure 2 illustrates the Weibull wind speed density distribution for Hsinchu Buoy and Liuqiu Buoy. The graph indicates that wind speeds in the Kaohsiung offshore area are concentrated between 1 and 4 m/s, while those in the Hsinchu offshore area are concentrated between 2 and 14 m/s.

3.2. Fixed and Floating Wind Turbines

In addition to optimizing wind energy, the water depth of offshore wind farms is a critical factor to consider. In regions where the water depth falls within the range of 50 m, fixed wind turbines are deemed more suitable for installation. For instance, the Hsinchu Buoy serves as an illustration, featuring a water depth of 24 m. Conversely, in areas with water depths surpassing 50 m, such as the Liuqiu Buoy near Kaohsiung with a water depth of 78 m, floating offshore wind turbines are considered more appropriate. The choice between fixed and floating turbines is influenced by the specific water depth conditions, with each type offering distinct advantages based on the depth of the offshore site.

This study focused on assessing the offshore wind power potential by examining two fixed wind turbines and two floating wind turbines.

Table 3. Wind turbine characteristics.

Characteristics	SWT-3.6-130	SG 8.0-167 DD	HTW5.2-127	SG 6.0-154
Rated power	3600 kW	8000 kW	5200 kW	6000 kW
Cut-in wind speed	4.0 m/s	3.0 m/s	4.5 m/s	4.0 m/s
Rated wind speed	12.2 m/s	12.0 m/s	14.0 m/s	13.0 m/s
Cut-out wind speed	25.0 m/s	25.0 m/s	25.0 m/s	25.0 m/s
Diameter	130 m	167 m	127 m	154 m
Swept area	13,273 m2	21,904 m2	12,668 m2	18,627 m2

outlines the key characteristics of the selected turbines. The chosen fixed wind turbine models are the Siemens SWT-3.6-130 [37], a globally utilized model, and the SG 8.0-167 DD [38], which has been deployed in offshore wind power plants in Taiwan. The hub height is specified as 135 m for the SWT-3.6-130 and 119 m for the SG 8.0-167 DD.

Additionally, the selected floating wind turbine models are Hitachi’s HTW5.2-127 [39] and Siemens’ SG 6.0-154 [40]. The HTW5.2-127 is an operational offshore wind turbine demonstration model in Japan, while the SG 6.0-154 is a demonstration model used in the Hywind Scotland project in Norway, marking the world’s first commercial floating wind farm. The hub height is specified as 90 m for the HTW5.2-127 and 101 m for the SG 6.0-154.

In the context of wind turbines, the power delivered is typically represented through a power curve, which establishes the relationship between wind speed and power output [41]. The power curves of the four wind turbines, including the manufacturer’s power curve, are illustrated in Figure 3.

3.3. Wind Energy Conversion

In wind turbine aerodynamics theory, the force F generated by the wind turbine blade, as proposed by Yeh and Wang [42], is represented as follows:

$$F=\mathsf{\rho }{A}_T{v}_T\left(v_1-v_2\right)$$

(5)

where v₁ represents the wind speed entering the blade (m/s); v₂ represents the wind speed leaving the blade (m/s); A_T represents the swept area of the blade (m²); v_T represents the average wind speed at the blade position (m/s); and ρ represents air density (kg/m³).

The power of the wind turbine blade’s rotation axis can be expressed as follows:

$$P=F{v}_T=\rho A_T{v}_T^2\left(v_1-v_2\right)$$

(6)

The change in kinetic energy from entering the wind turbine blade to exiting is:

$$D_{EK}=\frac{1}{2}{M}_A\left(v_1^2-v_2^2\right)$$

(7)

where M_A represents the mass flow rate of the wind, and its value is equal to $\rho A_T{v}_T$.

When the energy loss during the operation of the wind turbine is negligible, the energy acquired by the wind turbine can be represented by D_EK. Equality between Equations (6) and (7) can be used to determine this:

$$A_T{v}_T^2\left(v_1-v_2\right)=\frac{1}{2}{A}_T{v}_T\left(v_1^2-v_2^2\right)$$

(8)

The above equation simplifies to

$$v_T=\frac{v_1+v_2}{2}$$

(9)

Substituting v_T back into Equation (6), we get:

$$P=\frac{1}{4}\rho A_T\left(v_1^2-v_2^2\right)\left(v_1+v_2\right)$$

(10)

Assuming that the wind speed entering the blade v₁ is the same as the wind speed v, differentiating P with respect to the wind speed leaving the blade v₂ yields:

$$\frac{dP}{d{v}_2}=\frac{1}{4}\rho A_T\left(v+v_2\right)\left(v-3v_2\right)$$

(11)

When v₂ = v/3, this maximizes the wind energy captured by the blade. Substituting this into Equation (10), we get:

$$P_{max}=\frac{8}{27}\rho A_T{v}^3=\frac{16}{27}\frac{1}{2}\rho A_T{v}^3=C_{P_{max}}\frac{1}{2}\rho A_T{v}^3$$

(12)

where C_p represents the wind turbine blade efficiency. In ideal conditions [43], $C_{P_{max}}$ is approximately 0.593. However, in practical environments, wind turbines are influenced by factors such as air flow, leading to losses. We estimated the value of $C_{P_{max}}$ for each turbine based on Table 3, considering the rated power, rated wind speed, swept area, and air density (1.225 kg/m³). The results are presented in

Table 4. Power generation of wind turbines under average wind speed in offshore areas.

Region	Hsinchu Offshore Area		Kaohsiung Offshore Area
Wind turbine model	SWT-3.6-130	SG 8.0-167 DD	HTW5.2-127	SG 6.0-154
$C_{P_{max}}$	0.244	0.345	0.244	0.239
Average wind speed at hub height	10.30 m/s	10.17 m/s	5.77 m/s	5.83 m/s
Annual power generation	19 × 106 kWh	42.7 × 106 kWh	3.1 × 106 kWh	4.7 × 106 kWh

, with values ranging approximately between 0.24 and 0.35.

This study aims to account for the wind speed at the height where the wind turbine faces the wind. The buoy station wind speed is converted from a lower height to a higher height [42,44] as follows

$$\frac{v\left(Z_2\right)}{v\left(Z_1\right)}={\left(\frac{Z_2}{Z_1}\right)}^{\alpha }$$

(13)

where Z₁ is the height of the observation point (in this study, the buoy observation point height is 3 m), Z₂ is the hub height, and α is the friction coefficient. Since the offshore wind turbine is located above the sea surface, the value of α is taken as 0.1.

This study calculated the annual power generation of the wind turbine by utilizing the swept area data from Table 3, estimating $C_{P_{max}}$ as presented in Table 4, and factoring in the average wind speed at the hub height in offshore areas. The calculated results for wind turbine power generation at Hsinchu Buoy and Liuqiu Buoy are detailed in Table 4. The results indicate that due to the higher average wind speed in the offshore area of Hsinchu compared to that of Kaohsiung, the generated power in the offshore area of Hsinchu is significantly greater than that in Kaohsiung.

4. Model Development and Evaluation

This section aims to establish wind speed prediction models for offshore areas in Hsinchu and Kaohsiung. To understand the potential correlation between various feature data and offshore wind speeds in these regions, correlation coefficients (r) were calculated for each feature data and the wind speed at the offshore buoy. In this study, a correlation coefficient threshold of |r| ≥ 0.3 was used as a filtering criterion to exclude attributes with low correlation. Figure 4 illustrates the correlation analysis for attribute data from ground stations and buoys (a total of 13 attributes). The red lines in the figure represent |r| = 0.3. As shown in the figure, Hsinchu Buoy filtered out six attributes (Vm, Dm, Vg, WD01, WD03, WD05), while Liuqiu Buoy filtered out seven attributes (Vm, Vg, PS01, PS02, WD03, WD05, PP01).

This study divides the dataset into three periods: 2005 to 2016 (12 years) for training, 2017 to 2019 (3 years) for validation, and 2020 to 2022 (3 years) for testing. Additionally, the evaluation metrics employed in this study include root mean squared error (RMSE), relative RMSE (rRMSE), mean absolute error (MAE), and relative MAE (rMAE), calculated as follows:

$$\mathrm{MSE}=\sqrt{\frac{1}{N}{\displaystyle {{\displaystyle \sum }}_{i=1}^N}{\left(O_i^{Pre}-O_i^{Obs}\right)}^2}$$

(14)

$$\mathrm{r}\mathrm{RMSE}=\frac{\mathrm{RMSE}}{\frac{1}{N}{{\displaystyle \sum }}_{i=1}^N{O}_i^{Obs}}$$

(15)

$$\mathrm{MAE}=\frac{1}{N}{{\displaystyle \sum }}_{i=1}^N\left|\left(O_i^{Pre}-O_i^{Obs}\right)\right|$$

(16)

$$\mathrm{r}\mathrm{MAE}=\frac{\mathrm{MAE}}{\frac{1}{N}{{\displaystyle \sum }}_{i=1}^N{O}_i^{Obs}}$$

(17)

where N is the total number of data points, $O_i^{Pre}$ is the predicted value for the ith data point, and $O_i^{Obs}$ is the observed value for the ith data point.

4.1. Establishment of the Recurrent Neural Network Model

This study employs two different RNN models, LSTM and GRU, to establish wind speed prediction models for Hsinchu Buoy and Liuqiu Buoy. The forecasting target is the wind speed. First, an overview of the LSTM, GRU, and stacked LSTM models is provided below.

• In an LSTM model, each LSTM layer comprises recurrently connected memory blocks, as proposed by Wollmer et al. [45]. These blocks consist of self-connected memory cells and three multiplicative gate units: input gates, output gates, and forget gates. Hochreiter and Schmidhuber [27] introduced this design to tackle the vanishing gradient problem in traditional RNNs. The key innovation of LSTM is the inclusion of a memory cell, enabling the establishment of persistent long-term dependencies [46]. This addresses the challenge of vanishing gradients in conventional RNNs. The orchestrated functioning of gate units within memory cells allows LSTM models to store and access information over extended periods [47], facilitating effective learning from sequences with long-range dependencies. For a deeper understanding, refer to the seminal work of Hochreiter and Schmidhuber [27].
• The GRU model, a variant of the LSTM neural network, consists of input, forget, and output layers [48]. In this architecture, the input gate regulates information influx, the forget gate manages memory content through recurrent connections, and the output gate influences data used in calculating the output activation, directing information flow [49]. Similar to LSTM, GRU features gating units that modulate information flow without a separate memory cell [50]. This design effectively captures patterns and dependencies in sequential data. For deeper insights, see Cho et al. [51] and Chung et al. [52].
• The stacked LSTM architecture refers to an LSTM model consisting of multiple LSTM layers, also known as deep LSTMs. Similar to the RNN model framework, the stacked LSTM model comprises multiple LSTM layers stacked before being passed to a dropout layer and an output layer for the final output.

The input–output structure of the model is depicted in Figure 5, where data is sequentially fed into the model based on the time series. Each input comprises the current buoy wind speed, along with the buoy and ground attribute data filtered in the previous section. These models are configured to train on 12 target values (i.e., buoy wind speeds from t + 1 to t + 12). For each input data, with a time length of d (delay time), the model outputs the predicted buoy wind speeds for the next 12 h.

4.2. Parameter Calibration

The models in this study underwent parameter tuning using a trial-and-error approach. The tuned parameters for both LSTM and GRU models included the delay time, the number of hidden layers, and the number of neurons in each hidden layer. Additionally, the tuned parameters for the stacked LSTM model included the length of LSTM layers.

• For the parameter tuning of delay time in the models, testing starts with d = 1 h and progresses in 1-hour intervals up to a delay of 12 h. Figure 6 shows the RMSE errors under different delay times. The optimal delay time for the LSTM model at the Hsinchu Buoy remains at 10 h (Figure 6a), while for the GRU model, it ranges between 4 and 10 h (Figure 6b). For the Liuqiu Buoy, the LSTM model’s delay time ranges from 3 to 12 h (Figure 6c), and for the GRU model, it falls between 2 and 11 h (Figure 6d).

• For the number of hidden layers, the study explores 1 to 10 layers through trial and error. Figure 7 displays RMSE errors for different numbers of hidden layers. The optimal number of hidden layers for the LSTM model at the Hsinchu Buoy ranges between 1 and 2 layers, with 2 layers for t + 1 to t + 4 and 1 layer for t + 5 to t + 12. The optimal number of hidden layers for the GRU model is between 1 and 4 layers. For the Liuqiu Buoy, the optimal number of hidden layers for the LSTM model ranges between 2 and 5 layers, with 5 layers for t + 1 to t + 3 and t + 5, 3 layers for t + 4 and t + 6, and 2 layers for the rest. The optimal number of hidden layers for the GRU model ranges between 2 and 3 layers.

• The number of neurons in hidden layers significantly influences the model’s generalization. Too few neurons may lead to underfitting, while too many can cause overfitting. The study explores neuron counts from 10 to 150 in increments of 10. Figure 8 shows the results, indicating that for the LSTM model at Hsinchu Buoy, the optimal neuron count is 150, and for the GRU model, it ranges from 30 to 130. For Liuqiu Buoy, the LSTM model’s optimal range is 20 to 150, and the GRU model ranges from 130 to 150.

• Concerning the length of LSTM layers, the study investigates LSTM layers ranging from 1 to 10. Figure 9 illustrates the results, suggesting that for the stacked LSTM model at Hsinchu Buoy, the optimal number of LSTM layers falls within the range of 3 to 5. In the case of Liuqiu Buoy, the optimal range for the stacked LSTM model is identified as 2 to 5.

4.3. Test Results

After fine-tuning the parameters, the LSTM, GRU, and stacked LSTM models were implemented with optimized settings. Figure 10 and Figure 11 illustrate scatter diagrams depicting wind speed observations versus predictions at Hsinchu and Liuqiu Buoys for t + 1, t + 3, t + 6, and t + 12 from 2020 to 2022. The figures clearly show that as the prediction horizon extends, the predicted values tend to consistently underestimate the observed values.

Figure 12 assesses wind speed prediction models at Hsinchu Buoy and Liuqiu Buoy. In terms of RMSE, the LSTM, GRU, and stacked LSTM models at Liuqiu consistently outperform those at Hsinchu. Generally, stacked LSTM exhibits superior performance compared to LSTM and GRU at both locations. In the case of rRMSE, Hsinchu performs better overall, with stacked LSTM outperforming LSTM and GRU. Regarding MAE, Hsinchu’s LSTM excels at t + 1, while from t + 3 onward, Liuqiu’s LSTM, GRU, and stacked LSTM outperform Hsinchu. Stacked LSTM generally outperforms LSTM and GRU at both locations. In terms of rMAE, Hsinchu excels, with stacked LSTM outperforming LSTM and GRU.

5. Wind Rose Analysis and Seasonal Wind Speed Prediction Model

5.1. Seasonal Statistical Characteristics

Taiwan, situated between 120 to 122 degrees east longitude and 22 to 25 degrees north latitude, experiences a subtropical climate with distinct seasons. Summer is influenced by the Pacific maritime high-pressure system, leading to prevailing southwest monsoons. In winter, the dominance of the Siberian continental cold high-pressure system results in northeast monsoons. The seasons in Taiwan are spring (March to May), summer (June to August), autumn (September to November), and winter (December to February). Ganea et al. [11] suggested that incorporating seasonal patterns into wind speed prediction models enhances accuracy by capturing variations across the four seasons.

This study generates wind rose diagrams for Hsinchu Buoy and Liuqiu Buoy across the four seasons to observe the characteristics of wind speed and direction. Figure 13 illustrates the spring, summer, autumn, and winter wind rose diagrams for Hsinchu Buoy. The diagram shows that during summer, the prevailing wind comes from the southwest, aligning with the trend of southwest monsoons in Taiwan during this season. In spring, autumn, and winter, the prevailing wind comes from the northeast. Figure 14 displays the spring, summer, autumn, and winter wind rose diagrams for Liuqiu Buoy throughout the entire year. It reveals that during the summer season, Kaohsiung offshore experiences winds predominantly from the south, with wind sources spanning the entire southern region. The wind resources are also stronger during this season compared to others. In spring and autumn, the wind resources are similar, originating mainly from the northwest, while in winter, the wind resources are comparatively scarce.

5.2. Comparisons between Seasonal and Annual Models

This study, based on the seasonal characteristics discussed in the previous section, categorizes the seasonal model into summer mode and non-summer mode. In this context, the “summer model” refers to the predictive model established using wind speed data from June, July, and August, while the “non-summer model” corresponds to another predictive model constructed using data from the remaining months.

Figure 15 displays the RMSE evaluation metrics for seasonal models at the Hsinchu Buoy and Liuqiu Buoy, comparing them with the “annual model” (referring to the interannual model established in the previous chapter). By comparing the model errors between the annual and seasonal models, this evaluation assesses which type of model is more suitable for offshore Hsinchu (represented by the Hsinchu Buoy) and offshore Kaohsiung (represented by the Liuqiu Buoy).

• Figure 15a shows that the seasonal model offshore Hsinchu performs worse than the annual model before t + 6 in summer, but after t + 6, the seasonal model’s error becomes lower. Hence, the annual model can be chosen before t + 6, and the seasonal model after t + 6.
• Figure 15b illustrates the non-summer results offshore Hsinchu. With the exception of a slightly larger error in the GRU seasonal model, both the LSTM annual model and GRU annual model exhibit similar errors. In comparison, the stacked LSTM seasonal model demonstrates the smallest error.
• Figure 15c presents the summer results offshore Kaohsiung. In the LSTM models, the annual model shows higher testing errors than the seasonal model before t + 6, and this situation reverses after t + 6. In the GRU models, the difference in testing errors between the annual and seasonal models is relatively small. In comparison to LSTM and GRU, the stacked LSTM demonstrates the smallest error.
• Figure 15d illustrates non-summer results offshore Kaohsiung. Overall, the annual-based model exhibits larger errors, while the seasonal-based model demonstrates smaller errors. The stacked LSTM shows satisfactory results in both the annual-based and seasonal-based models.

6. Real-Time Power Generation Assessment

6.1. Annual Power Generation Forecast for Offshore Wind Turbines

In this section, the seasonal and annual models are used in conjunction with selected wind turbines for offshore Hsinchu and Kaohsiung. The predicted buoy wind speeds are converted to wind speeds at the turbine height, and the wind power potential for each model is calculated using the wind energy conversion formula. The operational efficiency is then assessed. For the offshore Hsinchu area, using the SWT-3.6-130 and SG 8.0-167 DD fixed offshore wind turbines and hourly observed wind speeds, the estimated annual power generation is 47.9 × 10⁶ kWh and 93.6 × 10⁶ kWh, respectively. In the offshore Kaohsiung area, with the HTW5.2-127 and SG 6.0-154 floating offshore wind turbines, the estimated annual power generation is 8.7 × 10⁶ kWh and 13.0 × 10⁶ kWh, respectively.

In the offshore Hsinchu area using the SWT-3.6-130 wind turbine, Figure 16a reveals that all models underestimate the annual power generation at t + 1, t + 3, t + 6, and t + 12. Overall, the stacked LSTM seasonal model is the closest to the observed values. At t + 1, t + 6, and t + 12, the GRU annual model performs worse than other models. When employing the SG 8.0-167 DD turbine in the offshore Hsinchu area (Figure 16b), overall, the stacked LSTM seasonal model achieves the highest annual power generation. At t + 1 and t + 12, the GRU annual model performs with the minimum annual power generation. At t + 3 and t + 6, the LSTM annual model exhibits the minimum annual power generation. Overall, for both the SWT-3.6-130 wind turbine and SG 8.0-167 DD turbine, the annual power generation from seasonal-based models is closer to the observed values.

In the offshore Kaohsiung area using the HTW5.2-127 wind turbine, Figure 16c indicates that, overall, the stacked LSTM seasonal model is the closest to the observed values. Subsequently, the annual power generation from the LSTM and GRU annual models is lower. When utilizing the SG 6.0-154 turbine in the offshore Kaohsiung area (Figure 16d), the predicted results from the models resemble those obtained using the HTW5.2-127 offshore wind turbine.

Furthermore, this study calculates the percentage of annual power generation for each model based on the observed data. Figure 17a,b shows that the percentage of annual power generation for offshore Hsinchu ranges approximately from 45% to 60% from t + 1 to t + 6, decreasing to around 40–45% at t + 12. Figure 17c,d demonstrates that the percentage of annual power generation for offshore Kaohsiung is roughly between 50% and 60% from t + 1 to t + 3, decreasing to 40–50% from t + 6 onwards.

6.2. Annual Power Generation per Unit Blade Swept Area

The energy output at Hsinchu and Kaohsiung offshore sites is influenced by the wind speed in the respective marine areas, the chosen offshore wind turbine model, and the depth of the sea. Hsinchu offshore employs fixed-type offshore wind turbines with higher tower heights and larger blade-swept areas. In contrast, Kaohsiung offshore, due to deeper waters, uses floating-type offshore wind turbines with lower tower heights and smaller blade-swept areas.

Figure 18 illustrates the evaluation of annual power generation per unit blade swept area using the stacked LSTM seasonal model for the Hsinchu and Kaohsiung offshore sites. It is evident that, for the Hsinchu offshore location, the utilization of SG 8.0-167 DD offshore wind turbines leads to a higher annual power generation per unit blade-swept area compared to SWT-3.6-130. In the Kaohsiung offshore area, the use of SG 6.0-154 offshore wind turbines results in slightly higher annual power generation per unit blade-swept area than HTW5.2-127. Additionally, the unit power generation per unit blade-swept area in Hsinchu offshore significantly surpasses that in Kaohsiung offshore.

6.3. Discussion

As different offshore wind turbines were used in the respective study areas, to ensure a fair evaluation and comparison, this study assumes the use of the floating wind turbine HTW5.2-127 and SG 6.0-154 in the offshore Hsinchu, allowing for a simultaneous comparison with SWT-3.6-130 and SG 8.0-167 DD. In the offshore Kaohsiung, fixed wind turbines SWT-3.6-130 and SG 8.0-167 DD were used (although fixed turbines may be less suitable due to deeper waters offshore Kaohsiung), enabling a comparison with HTW5.2-127 and SG 6.0-154.

Firstly, we conducted a comparison of the annual power generation for four sets of wind turbines at the offshore site near Hsinchu. According to the results in Figure 19a, we found that the SG 8.0-167 DD turbine exhibited the highest power generation, followed by SWT-3.6-130, SG 6.0-154, and HTW5.2-127, in descending order. Furthermore, in terms of annual power generation at the offshore site near Kaohsiung, as depicted in Figure 19b, the SG 8.0-167 DD turbine again showed the highest power generation, followed by SG 6.0-154, SWT-3.6-130, and HTW5.2-127, respectively. Notably, there were variations in power generation results between the two offshore locations, especially for SWT-3.6-130 and SG 6.0-154. To investigate the reasons for these differences, we explored factors such as the hub height of the turbines, which leads to variations in wind speeds at the same offshore location. Additionally, the swept area of different turbines also affects power generation. Therefore, based on our comparison, if the goal is to achieve maximum power generation, SG 8.0-167 DD turbines can be chosen for both Hsinchu and Kaohsiung offshore sites. However, considering the deeper water depth at the Kaohsiung offshore site, the option of a floating SG 6.0-154 turbine may be considered.

Taking into account the aforementioned considerations primarily centered on overall power generation, an exploration of the optimal solution for wind turbine deployment, especially in terms of blade size, can be undertaken. This exploration places emphasis on examining power generation per unit blade cross-sectional area. Figure 19c illustrates the annual power generation results for the four sets of turbines offshore near Hsinchu based on unit blade cross-sectional area. From the figure, it is evident that SG 8.0-167 DD remains the highest, while HTW5.2-127 is the lowest. In Figure 19d, the results for the four sets of turbines offshore near Kaohsiung are displayed, and the power generation trends mirror those offshore near Hsinchu. After a thorough comparison, it can be concluded that for both Hsinchu and Kaohsiung offshore sites, SG 8.0-167 DD is still the optimal choice in terms of power generation per unit blade cross-sectional area. However, when considering floating turbines for the Kaohsiung offshore site, due to the similar results for HTW5.2-127 and SG 6.0-154, one of them can be selected based on specific considerations.

7. Conclusions and Suggestions

7.1. Conclusions

The purpose of this study is to evaluate offshore wind energy potential, develop methods for predicting offshore wind power generation, and explore the potential of offshore wind energy in Taiwan. Machine learning techniques are employed to establish wind speed prediction models, and a method for real-time assessment of wind power generation is developed. The study focuses on two offshore regions in Taiwan, Hsinchu and Kaohsiung. The assessment of wind energy potential in these areas involves predicting the changes in offshore wind speeds for the next 12 h using buoy wind speeds combined with relevant data from ground stations. This enables the real-time estimation of future wind speed data and facilitates the assessment of the operational efficiency of offshore wind turbines.

In this study, machine learning methods, specifically long short-term memory (LSTM), gated recurrent unit (GRU), and stacked RNN with LSTM units neural networks, are employed to establish wind speed prediction models for the nearshore areas of Hsinchu and Kaohsiung. The wind speed prediction models are categorized into annual and seasonal modes based on seasonal characteristics. The annual mode utilizes yearly ground station and buoy data as input attributes, while the seasonal mode is further divided into summer and non-summer (other months). Summer is defined as June, July, and August.

This study establishes wind speed prediction models for the next 1 to 12 h in the offshore areas of Hsinchu and Kaohsiung. In the offshore region of Hsinchu, fixed offshore wind turbines, SWT-3.6-130, and SG 8.0-167 DD are used, while in the offshore region of Kaohsiung, floating offshore wind turbines, HTW5.2-127, and SG 6.0-154, are employed.

This study conducted a systematic analysis of four wind turbine configurations, encompassing aspects such as wind power potential assessment, wind speed prediction, and both fixed and floating wind turbine technologies. The research comprehensively considered the impact of various factors, including different offshore locations, turbine hub heights, swept areas, and wind field energy, on power generation. In conclusion, based on the research findings, it is recommended to choose the SG 8.0-167 DD turbine for offshore Hsinchu and the SG 6.0-154 turbine for offshore Kaohsiung. These recommendations serve as reference cases for selecting wind turbine configurations in the context of this study. Moreover, utilizing this evaluation process can also be extended to other regions worldwide for assessing the choice of offshore wind turbines.

In addition, we explored the global perspective on renewable energy trends and policy-making. Firstly, based on the report from REN21 in 2023 [53], the global installed capacity of renewable energy saw a significant increase, reaching a total of 348 GW in 2022, marking a 13% growth compared to the 306 GW added in 2021. Solar photovoltaic and wind power accounted for 92% of the new installed capacity, with solar contributing 70% and wind contributing 22%. According to the International Energy Agency (IEA), in a net-zero emission scenario, to achieve the 2030 targets set in the scenario assessment, the annual increase in renewable energy installed capacity needs to be more than 2.5 times the current rate. Specifically, to meet the IEA’s goals, the annual increase in wind power capacity needs to be boosted by 3.7 times.

In the offshore wind sector, six European countries and three Asian countries collectively added approximately 8.8 GW in 2022, bringing the global cumulative installed capacity for offshore wind power to 64.3 GW. However, the 2022 increase represents a 59% decrease compared to the previous year, primarily due to a slowdown in China’s offshore wind power development. Nevertheless, China has maintained its leadership in offshore wind power for the fifth consecutive year, contributing to over 58% of the new installed capacity, with the remaining share mostly distributed among Europe and Taiwan.

For a global perspective on renewable energy trends [54]:

• With the continuous growth of the global population, the demand for electricity is also on the rise, leading to a steady expansion of the global renewable energy market.
• Subsidies for renewable energy in advanced countries are trending towards stability or reduction, and countries are gradually introducing more competitive auction systems. Examining the current major global models for promoting renewable energy, these may include target setting, feed-in tariffs, electricity price differential subsidies, renewable energy ratios, green certificates, auction systems, and more.
• Local governments actively propose future development plans to expand the use of renewable energy in their regions. Globally, over 250 cities have committed to achieving a 100% renewable energy goal.
• The green supply chain is thriving, with companies increasingly demanding the procurement of renewable energy.

From a global perspective on policy-making: Internationally, the initial steps in developing renewable energy involve providing research and development subsidies and defining the scope of technology applications to reduce the application costs of renewable energy. Subsequently, measures such as equipment subsidies, tax credits, and financing incentives are implemented to alleviate the additional burden of investment. The enhancement of public awareness and support for government policies is achieved through demonstration projects and promotional campaigns. Finally, mechanisms like feed-in tariffs are employed to provide economic incentives through competitive rates for the purchase of renewable energy, thereby enhancing the market competitiveness of renewable energy. Alternatively, mandatory quota systems such as the renewable portfolio standard are established, imposing obligations on electricity providers to generate or procure a certain percentage of their energy from renewable sources, ensuring a stable demand for renewable energy in the market.

7.2. Limitations and Future Suggestions

This study focuses on conducting a systematic analysis of wind turbine configurations, encompassing aspects such as wind power potential assessment, wind speed prediction, and both fixed and floating wind turbine technologies. However, it is important to acknowledge that external factors, including government policies, technological advancements, and shifts in the energy market, could wield substantial influence on the wind power industry. Regrettably, these external factors were not extensively addressed in this study.

In future research, there is potential for further enhancing the performance of wind energy prediction models. One potential avenue for improvement is the consideration of ensemble models, which combine various types of models to enhance prediction accuracy. Additionally, a more comprehensive analysis of external factors, such as government policies, technological advancements, and fluctuations in the energy market, should be incorporated into the models to increase the predictability and resilience of the wind energy industry. This expanded investigation will contribute to a more holistic understanding of the dynamics influencing the wind power industry.

Second, during the installation of wind turbines, in addition to considering the aspects discussed in this study, there are many other factors that need to be taken into account, especially in terms of economic viability. These factors include electricity selling prices, the cost of power generation technology, government subsidies, and incentive measures. Therefore, it is recommended that future research delve deeper into the actual operational cost-benefit analysis of wind turbine systems. This will provide the industry with more competitive wind power generation solutions.

Third, the paper utilizes a stacked RNN with LSTM units for comparison with LSTM and GRU models. Following a rigorous model validation process and comprehensive testing, we present a comparative analysis that enhances the robustness of the validation for the selected models. However, for future considerations, we recommend incorporating additional emerging models for comparison, thereby enriching the scope of our research. This proactive approach will contribute to a more comprehensive understanding of the evolving landscape in neural network architectures.