As smart metering technology evolves, power suppliers can make low-cost, low-risk estimation of customer-side power consumption by analyzing energy demand data collected in real-time. With advances network infrastructure, smart sensors, and various monitoring technologies, a standardized energy metering infrastructure, called advanced metering infrastructure (AMI), has been introduced and deployed to urban households to allow them to develop efficient power generation plans. Compared to traditional stochastic approaches for time-series data analysis, deep-learning methods have shown superior accuracy on many prediction applications. Because smart meters and infrastructure monitors produce a series of measurements over time, a large amount of data is accumulated, creating a large data stream, which takes much time from data generation to deployment of deep-learning model training. In this paper, we propose an accelerated computing system that considers time-variant properties for accurate prediction of energy demand by processing the AMI stream data. The proposed system is a real-time training/inference system that deploys AMI data over a distributed edge cloud. It comprises two core components: an adaptive incremental learning solver and deep-learning acceleration with FPGA-GPU resource scheduling. An adaptive incremental learning scheme adjusts the batch/epoch in training iteration to reduce the time delay of the latest trained model, while trying to prevent biased-training due to the sub-optimal optimizer of incremental learning. In addition, a resource scheduling scheme manages various accelerator resources for accelerated deep-learning processing while minimizing the computational cost. The experimental results demonstrated that our method achieved good performance for adaptive batch size and epoch for incremental learning while guaranteeing a low inference error, a high model score, and queue stability with cost efficient processing.

No abstract

The data analysis platform used in smart grid is important to provide more accurate data validation and advanced power services. Recently, the researches based on deep neural network have been increasing in data analytic platforms to address various problems using artificial intelligence. The main problem to analyze multiple meter data based on deep learning is that the data distribution is varying according to both different client and time flow. Some studies, such as continual learning, are effective in dynamically fluctuating data distributions, but require additional complex computational procedures that make it difficult to construct an online learning system for processing data streams. In this paper, we proposed a hybrid deep learning scheduling algorithm to improve accuracy and accelerate learning performance in a multiple smart meter source environment, of which biased data feature varies dynamically. We use a simple analysis method, cosine similarity, to reduce computation complexity. By analyzing the frequency distribution of cosine similarity, a model recognizes that biased data feature of power consumption patterns. The skewed data distribution is reduced by using the zero skewness property of of an uniform distribution. The diversity of memory buffer was increased by update strategy which maximizes variance of pattern. When scheduling an online and offline gradient in different computational complexity, the proposed model reduces processing time by selectively calculating gradient considering the degree of data feature transition. To verify the performance of the proposed algorithm, we conducted three experiments with AMI stream data on the proposed method and the existing method of online learning. The experimental results demonstrate that our method can achieve reasonable performance in terms of trade-off between accuracy and processing time.

In the constrained computing environments such as mobile device or satellite on-board system, various computational factors of hardware resource can restrict the processing of deep learning (DL) services.Recent DL models such as satellite image analysis mainly require larger resource memory occupation for intermediate feature map footprint than the given memory specification of hardware resource and larger computational overhead (in FLOP) to meet service-level objective in the sense of hardware accelerator. As one of the solutions, we propose a new method of controlling the layer-wise channel pruning in a single-shot manner that can decide how much channels to prune in each layer by observing dataset once without full pretraining. To improve the robustness of the performance degradation, we also propose a layer-wise sensitivity and formulate the optimization problems for deciding layer-wise pruning ratio under target computational constraints. In the paper, the optimal conditions are theoretically derived, and the practical optimum searching schemes are proposed using the optimal conditions. On the empirical evaluation, the proposed methods show robustness on performance degradation, and present feasibility on DL serving under constrained computing environments by reducing memory occupation, providing acceleration effect and throughput improvement while keeping the accuracy performance.INDEX TERMS Single-shot pruning, channel pruning, lottery ticket hypothesis, DL model compression.

In some DL applications such as remote sensing, it is hard to obtain the high task performance (e.g. accuracy) using the DL model on image analysis due to the low resolution characteristics of the imagery. Accordingly, several studies attempted to provide visual explanations or apply the attention mechanism to enhance the reliability on the image analysis. However, there still remains structural complexity on obtaining a sophisticated visual explanation with such existing methods: 1) which layer will the visual explanation be extracted from, and 2) which layers the attention modules will be applied to. 3) Subsequently, in order to observe the aspects of visual explanations on such diverse episodes of applying attention modules individually, training cost inefficiency inevitably arises as it requires training the multiple models one by one in the conventional methods. In order to solve the problems, we propose a new scheme of mediating the visual explanations in a pixel-level recursively. Specifically, we propose DropAtt that generates multiple episodes pool by training only a single network once as an amortized model, which also shows stability on task performance regardless of layer-wise attention policy. From the multiple episodes pool generated by DropAtt, by quantitatively evaluating the explainability of each visual explanation and expanding the parts of explanations with high explainability recursively, our visual explanations mediatio scheme attempts to adjust how much to reflect each episodic layer-wise explanation for enforcing a dominant explainability of each candidate. On the empirical evaluation, our methods show their feasibility on enhancing the visual explainability by reducing average drop about 17% and enhancing the rate of increase in confidence 3%.INDEX TERMS Explainable AI (XAI), attention, class activation map (CAM), amortized model. I. INTRODUCTIONRecently, with the development of deep learning (DL) models, several studies [1]-[3] attempt to apply it on image analysis fields such as remote sensing or medical analysis. However, it is hard to clearly distinguish object classes on such satellite imagery due to its relatively low resolution, therefore, explanation on prediction is further required to provide reliability for the user via explainable AI (XAI) method [4], [5].