Accurately forecasting long-term future wind power is critical to achieve safe power grid integration. This problem is quite challenging due to wind power's high volatility and randomness. In this paper, we propose a novel time series forecasting method, namely Deep Conditional Generative Spatio-Temporal model (DCGST), and its high accuracy is achieved by tackling two critical issues simultaneously: a proper handling of the non-stationarity of multiple wind power time series, and a fine-grained modeling of their complicated yet dynamic spatio-temporal dependencies. Specifically, we first formally define theSpatio-Temporal Concept Drift(STCD) problem of wind power, and then we propose a novel deep conditional generative model to learn probabilistic distributions of future wind power values under STCD. Three different tailored neural networks are designed for distributions parameterization, including a graph-based prior network, an attention-based recognition network, and a stochastic seq2seq-based generation network. They are able to encode the dynamic spatio-temporal dependencies of multiple wind power time series and infer one-to-many mappings for future wind power generation. Compared to existing methods, DCGST can learn better spatio-temporal representations of wind power data and learn better uncertainties of data distribution to generate future values. Comprehensive experiments on real-world datasets including the largest public turbine-level wind power dataset verify the effectiveness, efficiency, generality and scalability of our method.

The study of time series forecasting is significant and useful in a variety of scenarios. However, due to the high degree of randomness and the complex contextual factors, it remains a difficult challenge. While several works based on machine learning and deep neural network have been proposed in recent years to address these challenges, most of them mine sequence features based on discrete points and overlook the fact that shape similarity plays an important role in inferring the future values. In this paper, we propose a seq2seq model with Shape Attention and Dilated Convolution (SADC) to tackle this problem. SADC contains two important phases: (1) Embedding with multi-scale dilated convolution. We first define the shape as a set of discrete points in a fixed-length window. The features hidden in the shape are then learned using multiple dilated convolutions with different kernels. (2) Inferring with shape attention. During this phase, we first present the shape attention, which aims to provide support information for inferring future values by generating the embedding vector of each prediction window based on shape similarity. The PreNet network is then built to predict the values using the embedding vector for each prediction window. The experimental results conducted on two datasets show that the performance of SADC model outperforms the state-of-the-art models on time series forecasting.