ML-Based Offshore Wind Forecasting Model

• Machine-learning-based offshore wind forecasting models are algorithmic systems that predict short-term wind speed and ramp events using advanced regression and ensemble techniques.
• They leverage robust data acquisition, signal processing methods like wavelet and empirical mode decomposition, and meticulous feature engineering to enhance prediction accuracy.
• Diverse algorithms including SVR, GBM, and deep learning are rigorously validated to support reliable grid integration and operational planning.

A machine-learning-based offshore wind forecasting model is an algorithmic system that predicts future wind conditions—including rapid “ramp” events—for offshore wind farms by leveraging historical and real-time measurements, advanced signal processing, and state-of-the-art regression or ensemble learning algorithms. The core objective is to provide highly accurate short-term forecasts of wind speed or power for operational planning, grid integration, and mitigation of risks associated with abrupt changes in wind production. The development and validation of these ML-based models draw from a combination of domain-specific data processing (e.g., wavelet decomposition), feature engineering, and rigorous comparative assessment against established statistical and physical benchmarks (Dhiman et al., 2020, Fischer et al., 2016, Walid et al., 25 Dec 2025).

1. Data Acquisition, Preprocessing, and Feature Design

ML-based offshore wind forecasting critically depends on the quality and structure of input data. Typical datasets originate from multiple offshore wind farms, with time-series wind speed observations (often at 10-minute intervals) and meteorological variables (pressure, temperature, humidity). For accurate hub-height prediction, wind speeds recorded at instrument height (e.g., 10 m) are adjusted to hub height (e.g., 90 m) using the logarithmic wind law:

$v_{h} = v_{10} \frac{\ln(h/z_0)}{\ln(10/z_0)},$

where $z_0=2\times10^{-4}$ m offshore.

Sophisticated feature extraction methods are routinely employed:

• Wavelet Transform (db4): Five-level discrete wavelet decomposition of normalized wind speed yields multiresolution features ($A_5$, $D_1 \ldots D_5$) that capture both low- and high-frequency variability, essential for ramp event characterization.
• Empirical Mode Decomposition (EMD): Ramp-power signals are decomposed into intrinsic mode functions (IMFs), and the log-energy entropy of each mode is computed to quantify intermittency(Dhiman et al., 2020).
• Engineered Statistics: Variance and mean of wind speed, lagged terms, rolling averages, and higher-order wind direction statistics are constructed for each farm (Fischer et al., 2016).

State-of-the-art studies emphasize extensive data cleaning (imputation, outlier removal), standardization, and careful selection of features via correlation analysis and variable importance ranking. Overly redundant or weakly correlated features are systematically discarded (Fang et al., 10 Mar 2025).

2. Model Architectures and Machine Learning Algorithms

Several classes of machine learning models are established for offshore wind ramp and power forecasting:

• Support Vector Regression (SVR) and Twin SVR (TSVR): SVR utilizes an $\varepsilon$-insensitive loss with Gaussian RBF kernel for nonlinear regression, while TSVR fits two related quadratic programs with the forecast as the average of two hyperplanes. TSVR consistently provides the lowest RMSE and ramp-event errors in offshore settings (Dhiman et al., 2020).
• Random Forest Regression (RFR): An ensemble method that averages over $T$ decision trees built on bootstrapped samples with random feature selection. RFR provides robust smoothing but tends to underfit extreme ramp changes.
• Gradient Boosted Machines (GBM): Stagewise additive models that iteratively fit weak learners to residuals, correcting errors over multiple passes. GBMs often match TSVR’s performance for large-scale ramp detection (Dhiman et al., 2020).
• CART-Bagging: Aggregation of full, unpruned regression trees with $B=500$ is shown to be the most stable and effective learning method for offshore wind power time series, outperforming parametric and logistic models (Fischer et al., 2016).
• LSTM and Deep Learning: Recent works introduce deep recurrent models (e.g., LSTM), sometimes with evolutionary hyperparameter tuning (CMA-ES), as in the CMAES-LSTM hybrid framework, providing state-of-the-art performance for 10-minute and one-hour horizons (Neshat et al., 2020).
• Hybrid Ensembles: Multi-model ensembles or direct multi-step global models (e.g., global CNN–RNN architectures) leverage spatial and temporal information more efficiently in the context of geographically distributed farms (Kazmi et al., 2023).

3. Model Training, Validation, and Hyperparameter Selection

Rigorous validation and optimization protocols are standard practice:

• Data Splitting and Rolling Validation: The data is split into 80% training and 20% test, often complemented by rolling time-series cross-validation without shuffling to preserve autocorrelation structure (Dhiman et al., 2020, Walid et al., 25 Dec 2025).
• Cross-Validation and Grid Search: Five-fold cross-validation on the training set is used to select optimal model hyperparameters, such as $C, \sigma, \varepsilon$ for SVR/TSVR, number of trees and depth for ensemble models, and learning rates for boosted trees and neural networks. Forecasting performance is evaluated on out-of-sample test data, using the model configuration with the minimal validation RMSE (Dhiman et al., 2020).
• Model Selection for Multistep Forecasts: Both recursive (iterated) and direct multi-output strategies for $k$-step-ahead forecasts are used, with direct models avoiding error accumulation in long horizons (Fischer et al., 2016).

4. Performance Metrics and Evaluation of Forecasting Skill

Model effectiveness is quantified using robust metrics:

• Root Mean Squared Error (RMSE):

$$\mathrm{RMSE} = \sqrt{\frac{1}{n}\sum_i(\hat y_i-y_i)^2}$$

• Mean Absolute Error (MAE):

$\mathrm{MAE} = \frac{1}{n}\sum_i|\hat y_i-y_i|$

• Mean Absolute Percentage Error (MAPE):

$$\mathrm{MAPE} = \frac{100}{n}\sum_i\left|\frac{\hat y_i-y_i}{y_i}\right|$$

For ramp event detection, absolute errors $R^{up}$ and $R^{down}$ are computed specifically at times when $|\Delta P_w|$ exceeds $\pm10\%$ of rated power.

Comparative benchmarks on northern European offshore sites yielded:

Model	RMSE (m/s)	$R^{up}$	$R^{down}$
TSVR	0.003–0.007	~0.003	~0.005
GBM	within 5–10% of TSVR RMSE	0.04–0.10	0.04–0.10
RFR	10–20% above GBM

SVR-variants, especially TSVR, exhibit pronounced generalization performance due to the $\varepsilon$-insensitive loss and effective nonlinear kernel mapping. GBMs deliver strong residual correction but show sensitivity to outlier ramp peaks (Dhiman et al., 2020).

5. End-to-End Forecasting Pipeline

The canonical pipeline for offshore ramp prediction, as outlined in (Dhiman et al., 2020), comprises:

1. Height Correction: Convert $v_{10}(t)$ to $v_h(t)$ via the logarithmic profile.
2. Feature Extraction: Apply five-level db4 wavelet decomposition; extract $A_5$, $D_1$–$D_5$ features.
3. Training/Test Split: Partition feature vectors and corresponding ramp targets for model fitting and validation.
4. Hyperparameter Tuning: Use 5-fold cross-validation to select the optimal configuration for each learning algorithm.
5. Model Fitting: Train the final model on all training data with selected hyperparameters.
6. Forecasting and Evaluation: Predict $\Delta\hat{P}_w(t)$ on test data, compute evaluation metrics (RMSE, MAE, MAPE, $R^{up}$, $R^{down}$).

for t in 1...T: v_h[t] = v10[t] * ln(90/z0) / ln(10/z0) A5, D1-D5 = db4_wavedecomp(v_h, levels=5) X = [A5, D1, ..., D5]; Y = DeltaP_w(t) split X, Y into train/test for M in {TSVR, RFR, GBM}: for each parameter set: cross-validate RMSE fit final model(s) on full training set DeltaP_w_hat = model.predict(X_test)

6. Model Selection, Interpretation, and Best Practices

• Selection of target variable is critical: predicting wind speed yields ~10% improvement in MAE/MAPE compared to direct power forecasting, as errors propagate nonlinearly through the cubic power law (Fang et al., 10 Mar 2025).
• Highly redundant or poorly correlated features (e.g., wind direction with low $|r|$ to speed at offshore sites) are excluded from the model to aid generalization.
• Model performance is robust to mild input noise and missing data: ensemble methods such as CART-bagging show limited error degradation when using spatially averaged (i.e., “deteriorated”) sensor input (Fischer et al., 2016).
• Simple model retraining and robust ensemble refresh schedules (weekly/monthly or on detected covariate drift) ensure operational stability and adaptability to changing conditions.

7. Applications, Impact, and Limitations

Accurately predicted offshore wind ramps directly aid grid operators in reserve planning and risk mitigation, minimizing grid stress and operational costs. For example, SVR- and GBM-based ramp forecasts are effective for real-time grid integration scenarios and for reducing wind curtailment risk (Dhiman et al., 2020). Fully integrated pipelines coupling ML with wavelet-based feature design provide a blueprint for industrial deployment in offshore environments characterized by high intermittency and complex sea–air dynamics.

Limitations include the need for high-quality, continuous wind speed and operational data and potential degradation in rare ramp event cases if such events are underrepresented. State-of-the-art models address these issues by incorporating both multiscale signal features and model ensembling approaches.

Overall, hybrid wavelet–machine-learning models combining SVR, GBM, and RFR variants, under a carefully engineered feature regime and robust validation protocol, represent the empirically established best practice for offshore wind ramp forecasting (Dhiman et al., 2020, Fischer et al., 2016).