Abstract:

Wind energy， a significant source of clean energy， has undergone rapid global development in recent years. However， wind power ramp events， caused by sudden changes in wind speed， adversely affect the power grid， potentially causing damage to wind turbines and leading to frequency instability. Existing research indicates that improving the prediction accuracy for these events is crucial for enhancing preventative capabilities. The low proportion of wind power ramp events within power data， coupled with a scarcity of corresponding samples， presents a significant challenge to improving prediction accuracy. Prediction methods for wind power ramp events are primarily categorized as either indirect or direct. Indirect prediction entails initially forecasting wind power， followed by the identification of ramp events. The drawback of this approach is that its predictive accuracy decreases significantly as the prediction horizon increases， making it unsuitable for long-term forecasting. Conversely， direct prediction first identifies historical wind power ramp events and then generates predictions. While the primary prediction targets are typically Boolean values， amplitude， and rate， existing models often fail to predict key information such as the start time and location of these events. A novel dual-meta pool prediction model is proposed capable of directly predicting the wind power ramp event vector. For data processing， the Hilbert curve is employed to map one-dimensional time series data onto a two-dimensional matrix， thereby effectively preserving positional information. A dataset expansion method based on random number generation is adopted to address the paucity of wind power ramp event data. Furthermore， a meta-data classification method based on time interval labels is proposed to reduce the dimensionality of the output data and the complexity of the prediction model. A two-stage classification model based on convolutional neural networks is proposed to establish the mapping between data and labels. For each tuple of meta-data， a corresponding predictive model for the meta-method is designed， and the mapping relationship between the meta-method pool and the meta-data pool is established by using the convolutional neural network with time interval labels. In contrast to traditional methods that employ a single prediction model for all data， the proposed method effectively diminishes the complexity of the neural network， enhances the relevance of data processing， and yields superior prediction accuracy. For the empirical analysis， output power data from three wind farms （SALTCRK1， DUNDWF3 and MEWF1） are randomly selected from Australian Energy Market Operator （AEMO） website （www.aemo.com.au）.The dataset covers the period from September 15， 2021， to June 17， 2023， with a 5-minute resolution and rated installed capacities of 54MW， 121MW， and 180MW， respectively. The empirical results for the proposed model show recall rates of 43.90%， 46.77%， and 43.12%， and mean absolute errors of 11.555， 23.861， and 24.558， respectively. Although the proposed dual-meta pool prediction model introduces an innovative approach， considerable scope remains for enhancing data processing techniques and predictive methodologies， necessitating further refinement through continued research efforts.

Key words: meta-data, method-data, neural network, wind power ramp