84 results about "Adaptive learning rate" patented technology

The invention relates to an electric energy quality prediction method based on similar days and improved LSTM. The method comprises the following steps: 1) collecting electric energy quality steady-state index data of a certain monitoring point in a certain region within a period of time and meteorological data of the region; 2) performing feature dimension reduction on the meteorological data byadopting a kernel principal component analysis method to obtain similar day feature vectors; 3) calculating meteorological factor matching coefficients of the historical day and the to-be-predicted day by adopting a grey correlation algorithm, and determining a similar day set; 4) selecting power quality historical data similar to the day to be predicted and similar day set data as a training sample set of the LSTM neural network, and optimizing LSTM neural network parameters by adopting a method of combining a Dropout algorithm and an Adam adaptive learning rate optimization algorithm, and taking the similar day feature vector as a model input variable to obtain a prediction result of the electric energy quality of each moment of the to-be-predicted day. Compared with the prior art, the method has the advantages of avoiding training overfitting, avoiding information interference and information repetition, improving prediction accuracy and the like.

The invention relates to a deep learning network intrusion detection method based on an improved learning rate. Determination method for learning rate and iteration frequency of improved deep belief network model in training is adopted, and the advantages of the improved deep belief network model and a softmax regression function are utilized to establish a deep belief network for network intrusion detection; and the softmax multi-classification combination model is used for training the model by using the public training data of the network intrusion detection data set and identifying and classifying the test data of the network intrusion detection data set by using the trained model. Rapid convergence of model parameters is realized by using an adaptive learning rate, and the optimized deep belief network-softmax multi-classification combination model is used for an intrusion detection system, the identification accuracy of attack behaviors can be effectively improved, and meanwhilethe detection efficiency can be improved.

The invention discloses an electric energy quality disturbance positioning and identifying method based on the lifting of the wavelet and the improvement of a BP neural network. The method comprises the following steps: using the Euclidean decomposition algorithm to obtain a db4 wavelet lifting scheme; performing lifting wavelet decomposition to the disturbance signal; in combination with the modulus maxima, performing the positioning detection to the disturbance mutation point peak; and utilizing the method of combining the self-adaptive learning rate with the incremental momentum term to improve and carry out disturbance identification to the BP neural network. The method of the invention can better obtain the disturbance period information, achieves fast positioning and high precision, and overcomes the shortcoming that a traditional BP neural network is prone to fall into local minimal points and slow convergence speed, therefore, achieving a high identifying rate for the power distribution network system's electric energy quality disturbance.

The invention discloses a decentralized federated learning method, device and system, and belongs to the field of federated learning, and the method comprises the steps: building a global communication network among a plurality of clients participating in federated learning, and enabling a communication path to exist between any two clients; each client respectively receiving the model parameters of other clients directly communicating with the client at the previous moment, respectively calculating the sum of the products of the obtained model parameters at the previous moment and the corresponding weight coefficients, and calculating the first product of the loss function gradient of the local model at the previous moment and a preset adaptive learning rate; updating a model parameter at the current moment to a difference value between the model parameter and the first product; and repeatedly executing the iteration updating operation until the loss function of the local model of each client is not higher than the corresponding threshold value, or until the number of times of repeated execution reaches the maximum number of iterations. While privacy and data security of each client are protected, each local model is globally trained.

The invention discloses a multi-view language recognition method based on unidirectional self-tagging auxiliary information. The method comprises the following steps: firstly, implementing self-tagging on current words and word-level auxiliary information by virtue of a tagging model, so that probability distribution of self-tagging auxiliary characteristics of the current words is obtained; then, decoding the probability distribution of the self-tagging auxiliary characteristics by virtue of Viterbi, so that relatively accurate auxiliary characteristics are obtained, and bidirectional auxiliary information is converted into unidirectional auxiliary information; and inputting the unidirectional auxiliary information, together with the current words, into a multi-view language model for analysis, so that accurate semantics of the current words can be obtained. The multi-view language recognition method provided by the invention has the characteristics that on the basis of the word-level auxiliary characteristics in a multi-view neural network, adverse influence on post-text information is eliminated, the various word-level auxiliary information is adopted, the word-level auxiliary characteristics, which are represented as a tree structure, are introduced to the multi-view language model for training, in the tagging model and the language model, stable operators are adopted to regulate various adaptive learning rates and the like.

The invention discloses a dual-layer long-short term memory (LSTM) network-based early-stage state identification method of a 10kV single-core cable. The LSTM network-based early-stage state identification method is applicably used in the electrical field. The LSTM network-based early-stage state identification method comprises the steps of firstly, selecting from observable electrical quantity toobtain five types of current observable electrical quantity, and extracting a time sequence from the five types of current observable electrical quantity to construct cable early-stage state combination time sequence characteristic matrix; secondly, constructing a dual-layer LSTM network of an time sequence handling input according to the characteristic of identification matrix size; and thirdly,performing model training under supervised learning by a self-adaptive learning rate optimization algorithm to obtain a cable early-stage state identification model. By the LSTM network-based early-stage state identification method, big mass running by the cable can be fully utilized, the time sequence is extracted from five types of observable data to construct a combined time sequence characteristic matrix as an input of the dual-layer LSTM network under the condition that statistical characteristic is not used, a corresponding relation between an input and an output is determined by handling capability of the dual-layer LSTM on the time sequence input, and the cable early-stage state identification is further completed. By the LSTM network-based early-stage state identification method,the identification accuracy can reach 99.06%.