Research On Methods For Power System Transient Stability Assessment And Wind Power Prediction Based On Deep Learning

Power system transient stability assessment and wind power prediction are two critical issues in ensuring security and improving the economics of power systems.The community has made tremendous efforts in those areas and scored remarkable achievements over several decades.Recently,the existing methods face severe challenges in terms of scope of application and accuracy of assessment or prediction since the new development of power system puts forward higher requirements for transient stability and wind power prediction.The development of deep learning has provided new possibilities for the breakthrough of the above two.The research on power system transient stability assessment and wind power prediction based on deep learning has become a timely and frontier topic in the field of power system research.This thesis is dedicated to in-depth research on those,and the major works and achievements of this thesis are as follows:A method for power system transient stability assessment based on a deep belief network is proposed.This method is able to discover hidden patterns of data and provide valuable information for transient stability assessment by utilizing a deep structure.The training of the assessment model involves not only supervised learning using labeled samples but also the unsupervised training using unlabeled samples,which is helpful for the improvement of the generalization ability.The results of the case study show that this method has advantages of high assessment accuracy,strong capability in handling irrelevant features,and short time-consuming for the time-domain simulations.A method for power system transient stability assessment based on a stacked autoencoder is proposed.This method breaks through the traditional two-stage assessment mode,and organically integrates “feature extraction” and “classification assessment” to achieve transient stability in an end-to-end manner.This method is free from the dependence on the hand-crafted features and is able to extract features from raw measurement data automatically,which ensures the information integrity to some extent.In addition,sparse constraints and Dropout techniques are applied in the training stage in order to alleviate the over-fitting of the model.The simulations show that the method is able to achieve transient stability assessment accurately.Moreover,fed with the measurement data of the phasor measurement units(PMUs),the assessment model can adapt to the real-life applications well.A method for power system transient stability assessment based on a deep sparse denoising autoencoder network is proposed.Taking into consideration different measurements with different physical nature,this method proposes a deep network with multi-branch structure.The features are extracted from multiple kinds of measurements by multiple sparse denoising autoencoders(branches)respectively and fused at the top of the model,and finally,output the assessment results.The simulations show that the method is able to effectively handle the multiple measurements simultaneously,and has the advantages of high accuracy and robustness in the noisy environment.Meanwhile,the assessment model has a more concise structure and lower computation complexity compared to the conventional fully connected networks,which has the potential of satisfying the requirements of the large-scale power systems.A method for multi-sites wind speed prediction using spatial correlation is proposed.The spatial correlation of wind speed is analyzed by visualizations intuitively,and a general mathematical description of the multi-sites wind speed prediction is provided.Based on the essential characteristics of the spatio-temporal sequence,a two-stage modeling strategy,i.e.,extracting spatial features firstly and then capturing the temporal dependencies,is proposed.Following the strategy,a wind speed prediction model based on convolutional neural network is constructed.The model is able to receive two-dimensional matrices directly,which avoids the loss of spatial information.The simulations show that the proposed method can effectively improve the prediction accuracy by using spatial correlation,and it is superior to the conventional machine-learning-based methods in terms of overall average performance and individual error control ability.A method for multi-sites and multi-step wind prediction using spatio-temporal correlation is proposed.The spatial and temporal correlations are analyzed from a mathematical point of view by using the sampling cross function.The problem of multi-sites and multi-step wind speed prediction is formulated as a spatio-temporal sequence prediction task,and a general mathematical description of the problem is given.The spatial and temporal models are used to extract the spatial features and establish the temporal dependencies of the wind speed spatio-temporal sequence respectively,so as to achieve the purpose of learning spatial and temporal correlations jointly.The simulations show that the proposed method can improve the prediction performance through the cooperation of the spatial and temporal models.Moreover,the scale of the model does not change with the expansion of the sequences,hence theoretically,the model can utilize historical sequence with arbitrary length and has good scalability.A method for wind power prediction based on multi-modal and multi-task learning is proposed.Wind power prediction using multi-source heterogeneous measurements is discussed,and a general mathematical description of the problem is provided.The prediction model is able to utilize multiple kinds of measurements simultaneously,which is mainly benefited from the multi-modal learning strategy.Multi-task learning strategy enables the model to integrate multiple wind power prediction tasks so as to achieve wind power sequence prediction.In order to effectively determine the hyper-parameters,orthogonal testing is employed for model selection,which significantly reduces the optimization space and saves computation resources.The simulations show that the proposed method is able to improve prediction accuracy by leveraging multiple measurements.Moreover,the mutual-information-based indicators for input variable selection may provide favorable guidance for further improvement of model performance.