With the increase of grid-connected photovoltaic (PV) power plants, measuring the uncertainty of PV output forecast due to multiple meteorological factors has become an important method to ensure the stable operation of the power system. However, the manual similar days' extraction methods of historical power features and cloud motion features, which are commonly used for PV spatial-temporal prediction, are inefficient and coarse-grained. In this paper, focusing on automatic attention of the local similar sequences in daily features, we improve the attention scoring method of the original transformer to increase the calculation accuracy of the similarity of PV power sequences. We construct an end-to-end PV power prediction framework using an encoder-decoder architecture. Based on the spatial-temporal similarity characteristics of PV output, the fine-grained temporal and cloud-covered spatial attentions embedded in the prediction framework are proposed using the improved similar sequence attention module. Among them, fine-grained temporal attention can dynamically position and utilize similar historical PV power sequences for the predicted period; cloud-coverage spatial attention automatically captures the neighbor site's matching PV power sequences. Experiments were conducted on a real PV dataset, and the root mean square error was reduced by 21.34% on average compared to the baseline model. The results proved the effectiveness of the proposed model.

It is agreed that portfolio selection models are of great importance for the financial market. In this paper, a constrained multi-period multi-objective portfolio model is established. This model introduces several constraints to reflect the trading restrictions and quantifies future security returns by fuzzy random variables to capture fuzzy and random uncertainties in the financial market. Meanwhile, it considers terminal wealth, Conditional Value at Risk (CVaR) and skewness as tri-criteria for decision making. Obviously, the proposed model is computationally challenging. This situation gets worse when investors are interested in a larger financial market since the data they need to analyze may constitute typical big data. Whereafter, a novel intelligent hybrid algorithm is devised to solve the presented model. In this algorithm, the uncertain objectives of the model are approximated by a Simulated Annealing Resilient Back Propagation (SARPROP) neural network which is trained on the data provided by fuzzy random simulation. An improved imperialist competitive algorithm, named IFMOICA, is designed to search the solution space. The intelligent hybrid algorithm is compared with the one obtained by combining NSGA-II, SARPROP neural network and fuzzy random simulation. The results demonstrate that the proposed algorithm significantly outperforms the compared one not only in the running time but also in the quality of obtained Pareto frontier. To improve the computational efficiency and handle the large scale securities data, the algorithm is parallelized using MPI. The conducted experiments illustrate that the parallel algorithm is scalable and can solve the model with the size of securities more than 400 in an acceptable time.

General-purpose Sparse Matrix-Matrix Multiplication (SpMM) is a fundamental kernel in scientific computing and deep learning. The emergence of new matrix computation units such as Tensor Cores (TCs) brings more opportunities for SpMM acceleration. However, in order to fully unleash the power of hardware performance, systematic optimization is required. In this paper, we propose Acc-SpMM, a high-performance SpMM library on TCs, with multiple optimizations, including data-affinity-based reordering, memory efficient compressed format, high-throughput pipeline, and adaptive sparsity-aware load balancing. In contrast to the state-of-the-art SpMM kernels on various NVIDIA GPU architectures with a diverse range of benchmark matrices, Acc-SpMM achieves significant performance improvements, on average 2.52x (up to 5.11x) speedup on RTX 4090, on average 1.91x (up to 4.68x) speedup on A800, and on average 1.58x (up to 3.60x) speedup on H100 over cuSPARSE.

In order to utilize the distributed characteristic of sensors, distributed machine learning has become the mainstream approach, but the different computing capability of sensors and network delays greatly influence the accuracy and the convergence rate of the machine learning model. Our paper describes a reasonable parameter communication optimization strategy to balance the training overhead and the communication overhead. We extend the fault tolerance of iterative-convergent machine learning algorithms and propose the Dynamic Finite Fault Tolerance (DFFT). Based on the DFFT, we implement a parameter communication optimization strategy for distributed machine learning, named Dynamic Synchronous Parallel Strategy (DSP), which uses the performance monitoring model to dynamically adjust the parameter synchronization strategy between worker nodes and the Parameter Server (PS). This strategy makes full use of the computing power of each sensor, ensures the accuracy of the machine learning model, and avoids the situation that the model training is disturbed by any tasks unrelated to the sensors.