Abstract:The prediction accuracy of short-term load forecast (STLF) depends on prediction model choice and feature selection result. In this paper, a novel random forest (RF)-based feature selection method for STLF is proposed. First, 243 related features were extracted from historical load data and the time information of prediction points to form the original feature set. Subsequently, the original feature set was used to train an RF as the original model. After the training process, the prediction error of the original model on the test set was recorded and the permutation importance (PI) value of each feature was obtained. Then, an improved sequential backward search method was used to select the optimal forecasting feature subset based on the PI value of each feature. Finally, the optimal forecasting feature subset was used to train a new RF model as the final prediction model. Experiments showed that the prediction accuracy of RF trained by the optimal forecasting feature subset was higher than that of the original model and comparative models based on support vector regression and artificial neural network.

Power quality signal feature selection is an effective method to improve the accuracy and efficiency of power quality (PQ) disturbance classification. In this paper, an entropy-importance (EnI)-based random forest (RF) model for PQ feature selection and disturbance classification is proposed. Firstly, 35 kinds of signal features extracted from S-transform (ST) with random noise are used as the original input feature vector of RF classifier to recognize 15 kinds of PQ signals with six kinds of complex disturbance. During the RF training process, the classification ability of different features is quantified by EnI. Secondly, without considering the features with zero EnI, the optimal perturbation feature subset is obtained by applying the sequential forward search (SFS) method which considers the classification accuracy and feature dimension. Then, the reconstructed RF classifier is applied to identify disturbances. According to the simulation results, the classification accuracy is higher than that of other classifiers, and the feature selection effect of the new approach is better than SFS and sequential backward search (SBS) without EnI. With the same feature subset, the new method can maintain a classification accuracy above 99.7% under the condition of 30 dB or above, and the accuracy under 20 dB is 96.8%.

Wind power combination probability prediction can effectively describe the uncertainty of wind farm output power and reduce the negative impact of this uncertainty on grid dispatching and operation. However, there is a conflict between coverage and interval width in probability prediction that affects the construction of the optimal prediction model. To resolve the conflicts between different probabilistic evaluation indicators, this paper proposes a new method for wind power combination probability prediction based on area gray correlation decision. First, the original wind power output data is reconstructed using energy-optimized variational mode decomposition to reduce the randomness of the original wind power signal. Processed wind power output data is used to establish an input feature set containing 96-dimensional historical wind power output data. Then, different Gaussian process regression prediction models are established based on 10 covariance functions. The area gray correlation method is used to calculate the area gray correlation degree of the five evaluation indicators, and a comprehensive evaluation of multiple indicators is carried out to eliminate the conflict between indicators. Finally, the weights of the prediction results of the 10 GPR models are determined according to the area gray correlation closeness, and the prediction interval and mean are reconstructed by combining the calculation results of each model. Simulation results show that the model determined by the new method has a more reliable prediction interval and faster prediction speed and can therefore provide decision support for wind power probability prediction and for the safe and stable operation of wind power grid connections.

Collaborative filtering recommender systems are now popular both commercially and in the research community. However, they are vulnerable to manipulation by malicious users, where attackers inject into some fake user profiles in order to bias the recommendation results to their benefits. To solve the problem, a lots of methods have been proposed but mainly focus on identification the attacker at the individual level, i.e., to find the fake user one by one, while do not consider the similarity between attack users. In this paper, we present an algorithm to detect the attackers in group level. It works based on an effective algorithm for detecting individual malicious user and an effective clustering algorithm. More precisely, we cluster all users into group, and then find the group characters of attacked items, finally we find the attack user group. We test the algorithm on a benchmark dataset using four kinds of typical attack models, the results show that our solution is both efficient and effective, particularly in the popular attack model and the segment attack model, and the performance is significant in the segment attack model with large attack size.