As the output from wind power generation is intermittent in nature, making the wind power output “dependable” is critical for seamless integration of wind generation. One of the most favorable solutions is incorporating Energy Storage System (ESS) with wind farms to establish Wind-Energy Storage Hybrid System. Since it requires capital investment for ESS installation, it is important to estimate appropriate storage capacity and charging/discharging rate of ESS for desired applications. In this paper, the fluctuation feature of wind power output is analyzed both in time domain and frequency domain. The degree of fluctuation is extracted and illustrated as Quantization Index (QI). Based on QI clustering, the wind scenario with largest power fluctuation is selected as “worst performance”, according to which, scheduling time horizon, along with the capacity and charging/discharging power of ESS, can be determined. After case study, the proposed model is proved to improve the generation scheduling process.

The precision and reliability of distributed photovoltaic (PV) power forecasting play a crucial role in commercial PV plants for power integration and operation. However, the stochastic and intermittent nature of solar radiation affects forecasting accuracy. What's more, the diverse location and various meteorological information have brought challenges to real-time distributed forecasting. Therefore, it is more difficult for forecasting distributed power than centralized power. To solve this problem, we propose a four-stage space-time hybrid method for power forecasting of distributed PV plants. Firstly, Global and Local Feature Fusion Network (GLFFNet) is proposed to forecast the power output of centralized PV plants. Secondly, the weather clustering type is obtained to get ready for studying the relationship between centralized and distributed PV plants. Then, uncovered coefficients are figured out to show the distribution features of PV power both in the time and space domain. Finally, the correlation models are established based on Copula analysis. According to the above steps, the centralized forecasting results are transformed to obtain the distributed forecasting model. The case study shows that the proposed GLFFNet is more suitable for 1-hour ahead forecasting. Compared to 9 different neural models, MAPE reductions of GLFFNet model (time steps of 1 hour and 15 minutes) are 12.17%-23.23% and 3.18%-18.22%, respectively. Thus, MAPE of the case distributed PV power forecasting models are 0.62% (Type A), 0.96% (Type B) and 3.21% (Type C), respectively.

Due to the fluctuation and randomness of photovoltaic power over time, accurate and reliable ultra-short-term photovoltaic power forecasting is significant for real-time dispatch and frequency regulation of power grids. In this paper, the improved BO-LSTM forecasting frame considering frequency correlation mapping is proposed. Firstly, the features of photovoltaic power are extracted and resolved according to power series frequency segments. Then, the established BO-LSTM forecasting model is adjusted based on the above extracted features in separate segment, and the results of deterministic forecasting are obtained. Furthermore, in order to obtain the reliable performance, the time-correlation algorithm is employed into the above deterministic forecasting model, which offers the base for probabilistic power forecasting. Finally, the above algorithms and forecasting framework are applied to the measurement data from a commercial photovoltaic power station in North China. Compared to the benchmark models, the Power Interval Normalized Average Width (PINAW) error of the proposed ultra-short-term forecasting algorithm has shown satisfied improvements. The PINAW has reduced by 8.4% (v.s. Adam-LSTM), 48.9% (v.s. Sgd-LSTM), 52.8% (v.s. Adagrad-LSTM), 9.1% (v.s. Rmsprop-LSTM), 97.2% (v.s. Adadelta-LSTM), 86.8% (v.s. Adam-mlp), 87.4% (v.s. Sgd-mlp), 90.9% (v.s. Adagrad-mlp), 86.5% (v.s. Rmsprop-mlp), and 99.7% (v.s. Adadelta-mlp).

The distributed photovoltaic (PV) power stations within the entire county exist spatiotemporal correlation. Merely considering temporal correlation makes it challenging to meet the day-ahead scheduling demands. This paper proposes a distributed PV county-level day-ahead power prediction method based on grey relational analysis and the Transformer-Graph Convolutional Attention Network (Transformer-GCAN) model. Firstly, the grey relational degree is used to measure the relevance between each distributed PV stations, and the connection relationship of the station graph is determined based on the analysis results. Secondly, the Transformer network is utilized to extract the temporal features of each PV sequence in the graph. Based on the Graph Convolutional Network (GCN) model, a Graph Attention Mechanism (GAT) is introduced to dynamically extract spatial features between each photovoltaic station in the graph. Finally, the integration of spatiotemporal features is achieved through a fully connected neural network, enabling day-ahead power prediction at the county level. Case analysis results demonstrate that compared with the Transformer-GCN model, the Root Mean Square Error (RMSE) of the power prediction model proposed in this paper is reduced by 11.90%, 15.72% and 19.61% respectively in sunny days, cloudy days and rainy days.