Wind Power Prediction across Different Locations using Deep Domain Adaptive Learning

A substantial rise in global energy demand is observed due to the growing population, rapid industrialization, and progress in technology. While fossil fuel sources are being depleted to match this increasing demand, they also have adverse effect on our environment. Thus, renewable energy has become an effective solution to this energy crisis. Only in 2023, the renewables installed over the world was 510 GW as reported by the International Energy Agency (IEA) [1]. Due to the advancement of technology and lowering costs, wind energy has become one of the prevailing renewable sources. Currently, the global installed wind energy capacity is more than 710 GW [1]. This is an increase by a factor of almost 100 times over the past two decades [2].

Similar to most other renewable resources, wind energy production is highly intermittent and dependent on weather conditions. The fluctuations in wind power can lead to several issues, such as frequency deviation, harmonics, inter-area oscillations, economic dispatch problems within electricity markets, etc. [3]. Therefore, reliable prediction of wind power generation is of colossal importance. However, as wind power is dependent upon weather conditions, the generation varies widely over regions even for similar wind farm setup. Weather variables are location-dependent and deviate vastly from one climatic region to another, consequently causing deviations in wind power generation profile. Wind power generation can be predicted with statistical and machine learning (ML) methods, however, the prediction model remains mostly location-specific. The drawback of location-specific prediction models is that each model needs to be trained individually, requiring a large amount of data. However, reliable data collection in bulk quantities can be a challenging task in some locations due to adverse conditions. In such cases, domain adaptation or transfer learning methods can help immensely. Domain adaptation allows an ML model trained on a source dataset to predict the outcomes of another dataset called the target dataset with a comparable but different probability distribution. Domain adaptation makes the model more robust and computationally more efficient as it becomes more location-agnostic and requires less data from the test dataset. The motivation behind our work revolves around addressing this issue.

The contributions of this work can be summarized as:

1. 1.

   Proposing a domain adaptation-based location-agnostic model to predict wind power generation that achieves both higher accuracy and faster convergence.

2. 2.

   Generating a novel dataset that combines features from two distinct datasets and identifying relevant features through the application of random forest.

Wind power forecasting has been extensively researched in the literature. Most of the forecasting methods can be broadly classified into three approaches. They are (i) physical approach, (ii) statistical approach, and (iii) machine learning and deep learning-based approach. The physical approach has been investigated for wind power forecasting in several papers [5, 6, 7]. Statistical methods have also been studied in literature [8, 9, 10]. However, more recently ML and deep learning-based methods have gained popularity.

Transfer learning and domain adaptation-based deep learning models have also gained interest in in contemporary literature. Transfer learning is used with auto-encoding to achieve better prediction results for wind forecasting [11, 12]. In [12], auto encoding and transfer learning have been applied to determine the patterns in homogeneous and heterogeneous characteristics of wind turbine dynamics for wind power prediction. The paper claims that heterogeneous characteristics can stem from differences in operating conditions, locations, and operational strategies, etc. and compares the results of different methods with their proposed method for two wind farms. The proposed method outperforms the other methods in terms of both accuracy and time. However, the reason behind the varying heterogeneous characteristics of these two wind farms is absent in this work. A deep neural network (DNN)-based architecture is proposed with transfer learning application [11]. The goal is to predict the wind speed in newly built wind farms, lacking data, exploiting the data from previously built wind farms, where sufficient data is present. The data from four wind farms in China is used here. However, they are not located far enough to show enough variability in climatic conditions. A novel serio-parallel architecture is proposed in [13], where several parallel Convolutional Neural Networks (CNN) are separately connected to a Long Short Term Memory (LSTM) network. Transfer learning is applied to this network by adopting the Adam optimization to retrain the frozen layers of the pre-trained feature extractor and the fully connected layers. The dataset is comprised of six separate wind farms although they belong to the same climatic region of Inner Mongolia, China. Model-based transfer learning is applied in [14] to construct a Multi-layer Extreme Learning Machine (MLELM). Output parameters of MLELM are further optimized by using particle swarm optimization. Instead of providing the expected value of wind power, the proposed method provides optimized probability quantiles of the predicted values. However, all eight wind farm data comes from the same province in China. Transfer learning along with gradient boosting method also has been explored for some cases [15, 16]. An instance-based transfer learning embedded with gradient-boosting decision trees is proposed in [16]. Similar to [14], this paper also tries to find out the probability quantiles of the predicted wind power. An optimization method is proposed for assigning weights to the auxiliary training sets based on the relatedness of source and target tasks. However, all the ten wind farms used here are from Australia. Two transfer learning approaches Multi-task Learning (MTL) and Bayesian Embedded Multi-task Learning (BE-MTL) are proposed and compared with other regular ML methods namely, Random Forests, Gradient Boosted Regression Trees, and simple Multi-Layer Perceptrons [17]. The MTL and BE-MTL methods show superior performance while working with less available data. However, all 19 wind farm data are from southwest Germany. In [18], instead of using data from a single wind farm as a source domain, a cluster of wind farm data is used as multiple source domains. This multi-source domain adaptation (MSDA) uses different weights for each source domain with different but comparable probability distributions. This method shows a significant reduction in prediction error. Nonetheless, the work focuses on wind farms with similar weather patterns and thus uses wind stations that are closely located. Recently [4, 19] has shown the application of domain adaptation in solar power and rain precipitation prediction, respectively in a location-agnostic manner.

As we can see from above, there are multiple research works that focused on the use of transfer learning or domain adaptation to improve the forecasting results of wind power. Yet these works have not leveraged the fact that a pre-trained model of one climatic region’s wind farm can be used to predict the wind power generation of other climatic regions through domain adaptation. Therefore, our work centers around this issue and tries to generate a location-agnostic domain adaptive model.

In this paper, we propose a method to predict the wind power generation in a region at a certain time, based on the weather and climatological characteristics of the region. A model that was once trained using the data of one country will be used further to adapt to the domain of a different country. The first country where the model was actually trained is referred as the source domain, and the other country where we intend to adapt the model is referred as the target domain. In the source domain, the model is trained in a fully supervised manner. For this work, we assume that the target domain does not have any data available, i.e. the only thing it has is a pre-trained model from the source domain. Then the pre-trained model is adapted for the target domain by only updating the last few layers of the pre-trained model. Therefore, the overall training procedure becomes faster and the computational burden is reduced for larger datasets. In short, our work follows two main steps, i.e., train the model in the source domain, followed by adapting the model in the target domain. Figure 1 illustrates the overview of the overall approach.

An original dataset is created for domain adaptive wind power prediction. The creation of this novel dataset is executed through incorporation of appropriate data from two different datasets; one containing the hourly wind generation data and another containing the meteorological data. The wind generation data is collected from the EMHIRES dataset [20]. This dataset contains 30 years’ worth of hourly wind generation data for 36 of the European countries. For each country, the wind generation is normalized relative to the installed wind generation capacity. For this work, the latest available five years’ worth of data has been used. Hourly meteorological data is retrieved from [21] for the corresponding wind power generation period. The final dataset is created by merging the two datasets according to their corresponding hourly slots. Ultimately, the meteorological dataset contains 21 weather parameters shown in Table I. However, only a handful of the weather features are useful for wind power prediction as inclusion of irrelevant features can make the training process slower and may even compromise the fidelity of the forecast model.

Random forest algorithm is considered as an effective tool for the selection of compatible features [22]. Following the implementation of this algorithm, only six effective features were selected for prediction namely, temperature, dew point, snow, snow depth, wind speed, and cloud cover.From the dataset, we have analyzed the correlation of data for wind power generation. Figure 3 displays the correlation among the selected six features related to wind power generation through a heat map. For our experiments, we choose three different countries Germany, France, and the United Kingdom. The location of these three countries is such that they have a comparable yet distinguishable weather profile. The dataset comprises wind data and various weather features from three distinct European countries, effectively capturing the spatial disparities in data distribution across these regions. However, while the dataset acknowledges these differences, it’s our proposed model that solves the challenge of properly addressing them.

In this work a source-free and location-agnostic domain adaptive model was presented for wind power generation prediction. The proposed model gets pre-trained on one country’s data (source domain) and this model is then used to predict the wind generation of another country (target domain) using selective weather features. The proposed approach performs significantly better than the traditional approach. The assumption of source data being unavailable while training the target domain makes it less data-intensive and computationally more efficient. Furthermore, two additional studies are carried out to investigate the impact of updating the partial network during target domain training and the impact of using selective features only, exhibiting the efficacy of our design. While our current implementation leverages supervised learning, we plan to explore unsupervised learning techniques in the future for further extension and refinement of our approach. Additionally, we aim to incorporate federated learning methodologies, bolstered by robust security features, to enhance the privacy and integrity of our model across distributed environments.