Heating load forecasting is the premise for guiding heating operation management and dispatching. Heating load forecasting is a time series prediction problem which requires us to predict the real-time heating loads in the next 24 hours using available historical records and weather information. In this paper, we propose a model for short-term heating load forecasting based on a properly designed strand-based long short term memory (LSTM) recurrent neural network. We present how the data are pre-processed, and the loss function is designed to improve the model's performance. Furthermore, an ensemble strategy is incorporated with the LSTM model to enhance its generalization and robustness. On offline (historical) testing data, the proposed model performs satisfactory predictions which meet the requirements of the local power plant. In addition to offline tests, we also implement the model to an online system of a power plant in Shandong province, China. The model made continuous forecasting without human interference for four months during the heating season of 2018. The model reported satisfactory online testing results that were comparable with the offline experiments using historical data.
11 October 2018 in IOP Conference Series: Earth and Environmental Science
We propose a long short-term memory (LSTM) network based encoder-decoder (E-D) model for wind power prediction (WPP). The LSTM-based E-D model is constructed as an auto-encoder for mapping the wind power (WP) time-series into a fixed-length representation, state of the trained E-D LSTM. Then, the representation concatenated with weather forecasting information is used as a new input to another multiple LSTM network to make WPP. Real data collected from a wind farm with capacity of 50 MW of Shan Xi province were used to verify the conclusions. Results illustrate that the proposed method improves the model generalization ability and lowers misspecification risk by utilizing the WP time relationship through autoencoding (AE) process. Combining extracted representation with weather forecasting information further improves the prediction accuracy.
The collaboration between renewable energy (RE) and an electric vehicles battery switch station (BSS) is a win-win strategy for both. In order to effectively coordinate and manage RE and BSS belonging to different investors, a new cooperative operation mode considering a multi-stakeholder scenario is proposed. By analyzing the relationship of cooperation and competition between them, their pursuit of maximizing their own interests is modeled as a Stackelberg game model. As the leader, the RE company determines the charge/discharge prices to guide the demand response of BSS. As the follower, the BSS optimizes its charge/discharge strategy according to the prices to maximize its own profits. The proposed model no longer requires that the RE company and BSS belong to a same interest subject, and there is no need for unified coordinated control between them. To solve the game equilibrium, a solution combining the backward-induction method and genetic algorithm is proposed by comparing and analyzing the characteristics of the strong and weak Stackelberg equilibriums. Finally, the effectiveness of the proposed model was validated by case studies. The simulation results show that both the RE company and the BSS have the motivation to participate in the game, and that a win-win outcome can be achieved automatically in the process of pursuing their own interest maximization in the game.
3 August 2018 in IOP Conference Series: Materials Science and Engineering
A gated recurrent units (GRU) network based encoder-decoder (E-D) model is proposed for combined heat and power (CHP) heat load forecasting. First, we use the GRU based E-D model as an auto-encoder to map the historical CHP heat load time series into a fixed length representation. Then, concatenate the representation with weather forecasting information as a new input to another multiple GRUs network for heat load prediction. Data collected from Rizhao, Shandong province is used to verify the conclusions. Results illustrate that GRU-based E-D model can give a more accurate forecast to the short-term CHP heat load compared with non-auto-encoded RNN model.
Power system ramping demand consists of deterministic and uncertain components. This study focuses on predicting uncertain ramping demand influenced by meteorological factors. First, an adaptive trend identification algorithm is proposed to extract key features of upward and downward ramps, including magnitude and duration. Multidimensional correlation analysis quantifies the relationships among net load characteristics, historical load forecasts, weather predictions, historical prediction errors, and uncertain ramping demand. The minimum redundancy maximum relevance (mRMR) method is applied to compress input data features, and a TimeXer-modified prediction framework is developed to capture long-term dependencies in time-series data while integrating meteorological features through attention mechanisms, outputting confidence intervals for uncertain ramping demand. To enhance performance under extreme weather conditions, a method for defining extreme weather indicators is proposed, establishing thresholds based on historical meteorological data distributions and creating binary features for each time step to denote extreme weather. The original TimeXer model’s output layer is adapted to predict both the mean and variance of ramping demand, addressing uncertainty in extreme weather scenarios. A Gaussian negative log-likelihood loss function is employed for training to optimize uncertainty prediction performance. Validation using 2024 operational data from Shandong, China, indicates that the proposed method improves prediction accuracy by 18.4%, achieves a 92.5% interval coverage rate, and produces narrower prediction interval bandwidths under equivalent coverage rates, effectively revealing the impact of meteorological factors on ramping demand, especially under extreme weather conditions, where uncertainty prediction is notably more accurate.
A multi-layer LSTM (Long short-term memory) model is proposed for condenser vacuum degree prediction of power plants. Firstly, Min-max normalization is used to pre-process the input data. Then, the model proposes the two-layer LSTM architecture to identify the time series pattern effectively. ADAM(Adaptive moment)optimizer is selected to find the optimum parameters for the model during training. Under the proposed forecasting framework, experiments illustrates that the two-layer LSTM model can give a more accurate forecast to the condenser vacuum degree compared with other simple RNN (Recurrent Neural Network) and one-layer LSTM model.