88 results about "Model aggregation" patented technology

The invention provides a federal learning method and device, and the method comprises the steps: adding a verification round in a federal learning process, enabling a server to transmit a model updating parameter returned by a to-be-verified client last time to an auxiliary client, and carrying out the training through the local data of the auxiliary client, and calculating the deviation between the loss value when training of each auxiliary client in the verification round is terminated and the previous round, and if the number of the auxiliary clients with the deviation greater than a set threshold is greater than a set proportion, marking the client to be verified as an abnormal client. According to the method, the abnormal client can be quickly and effectively identified under the condition that each client is not informed, and furthermore, the weight in the model aggregation process is adjusted according to the deviation in each verification round corresponding to the abnormal client, so that the adverse effect of the abnormal client on global model updating is prevented.

The invention provides a virtual power plant day-ahead scheduling optimization model. A model aggregation unit comprises a gas turbine, a wind turbine generator set, a photovoltaic set, a water drawing energy storage power station and loads. For the characteristics that the electricity price probability distribution description is relatively accurate and the prediction is relatively high, random programming is adopted to process the uncertainty of the electricity price; and for the characteristics that the wind power and photovoltaic output probability distribution is difficult to precise describe and the prediction precision is relative low, an information gap decision theory (IGDT) is adopted to process the uncertainty of wind power and photovoltaic output, different weights are provided to wind power and photovoltaic output deviation coefficients, and the IGDT is enabled to simultaneously process the uncertainty of wind power and photovoltaic output. In addition, for the blindness of uncertainty decisions and the different risk degrees of different strategies, the risk cost is introduced, and the risks corresponding to different decision schemes are quantified. According to the invention, a larger decision making space is provided for a decision maker, and the VPP is enabled to make the optimal decision under more conditions, so that the benefit of the virtual power plant (VPP) is increased.

The invention discloses a decentralized federated learning framework based on heterogeneous computing power perception. The decentralized federated learning framework comprises a cloud coordinator and a plurality of equipment ends. The cloud coordinator is used for management, training and parameter updating scheme generation during operation and regular model backup. The equipment end is used for transmitting equipment information to the cloud coordinator, operating a model locally and updating equipment end parameters. The cloud coordinator acquires the least common multiple of the one-time training time of the equipment end as a super cycle, the equipment end calculates different step lengths in the super cycle, and the model is aggregated in the integral multiple of the super cycle. Different local steps are operated according to different computing capabilities of equipment, and in the model aggregation process, negative effects of slow nodes are reduced; and a distributed point-to-point communication mode is adopted, and the communication pressure of the central server in the distributed training process can be eliminated under the condition that the overall communication traffic is not increased.

The invention discloses a federated learning method with high communication efficiency in a wireless communication scene. The method comprises the following steps: S1, constructing a federated learning framework from three aspects of a federated learning system, a training algorithm and a communication model in the wireless communication scene; s2, aiming at the constructed federal learning framework, carrying out convergence analysis on the training process of the federal learning framework; and S3, constructing an optimization problem about the federated learning framework according to a convergence analysis result, and solving the problem through a joint optimization method for equipment selection and beam forming. According to the method, on the basis of air calculation and a second-order training algorithm, on one hand, low-delay model aggregation is achieved through the waveform superposition characteristic of a channel, on the other hand, the number of iteration rounds needed by training is reduced through the rapid convergence characteristic of the second-order algorithm, and the communication bottleneck problem existing in most wireless federated learning methods at present is solved. Meanwhile, the training accuracy is further improved through the provided joint optimization method for the federated learning framework.