A significant amount of distributed photovoltaic (PV) generation is “invisible” to distribution system operators since it is behind the meter on customer premises and not directly monitored by the utility. The generation essentially adds an unknown varying negative demand to the system, which causes additional uncertainty in determining the total load. This uncertainty directly impacts system reliability, cold load pickup, load behavior modeling, and hence cost of operation. Thus, it is essential to create low-complexity localized models for estimating power generation from these invisible sites behind the meters. This article proposes an adaptive machine learning framework to: a) learn using weather data and a minimal number of BTM PV generation measurement sensors, b) forecast PV generation using weather, location of PV, and trained ML model at location for unmeasured BTM PV; c) use estimated PV and net load measured by smart meter or smart transformer to estimate total true load at each time step; and d) learn the specific load patterns eventually to adapt localized models. The proposed framework's core idea is to transform the data such that: a) the machine learning model can effectively utilize the time dependency of measurements; and b) the measurements are transformed into a lower dimensional space to reduce complexity while maintaining accuracy. The transformed measurements are then used to train the machine learning models for load/PV disaggregation. Machine learning models investigated include linear regression, decision tree, random forest (RF), and multilayer perceptron. The proposed framework's efficacy is demonstrated using two datasets, a real dataset from Hawaii and a simulated dataset using detailed models in GridLab-D. Several test/training split scenarios, including 90-10% split, one-month-out, one-season-out, and panel-independent split are presented to provide a thorough evaluation of the proposed framework. Results on both datasets show that the proposed framework can estimate PV generation with high accuracy using low-complexity methods. The accuracy results are comparable to higher complexity models (e.g., deep architectures), and RF is found to provide superior performance with these specific datasets compared to the other ML models investigated.
Human activity recognition (HAR) plays a central role in health-care, fitness and sport applications because of its potential to enable context-aware human monitoring. With the increase in popularity of wearable devices, we are witnessing a large influx in availability of human activity data. For effective analysis and interpretation of these heterogeneous and high-volume streaming data, we need powerful algorithms. In particular, there is a strong need for developing algorithms for robust classification of human activity data that specifically address challenges associated with dynamic environments (e.g. different users, signal heterogeneity). We use the term robust here in two, orthogonal senses: 1) leveraging related data in such a way that knowledge is transferred to a new context; and 2) actively reconfiguring machine learning algorithms such that they can be applied in a new context. In this paper, we propose an architecture that combines an active learning approach with a novel deep network. Our deep neural network exploits both Convolutional and Long Short-Term Memory (LSTM) layers in order to learn hierarchical representation of features and capture time dependencies from raw-data. The active learning process allows us to choose the best instances for fine-tuning the deep network to the new setting in which the system operates (i.e. a new subject). We demonstrate the efficacy of the architecture using real data of human activity. We show that the accuracy of activity recognition reaches over 90% by annotating less than 20% of unlabeled data.
30 October 2018 in IEEE Transactions on Mobile Computing
Machine learning algorithms, which form the core intelligence of wearables, traditionally deduce a computational model from a set of training data to detect events of interest. However, in the dynamic environment in which wearables operate, the accuracy of a computational model drops whenever changes in configuration or context of the system occur. In this paper, using transfer learning as an organizing principle, we propose a novel design framework to enable autonomous reconfiguration of wearable systems. More specifically, we focus on the cases where the specifications of sensor(s) or the subject vary compared to what is available in the training data. We develop two new algorithms for data mapping (the mapping is between the training data and the data for the current operating setting). The first data mapping algorithm combines effective methods for finding signal similarity with network-based clustering , while the second algorithm is based on finding signal motifs . The data mapping algorithms constitute the centerpiece of the transfer learning phase in our framework. We demonstrate the efficacy of the data mapping algorithms using two publicly available datasets on human activity recognition. We show that the data mapping algorithms are up to two orders of magnitude faster compared to a brute-force approach. We also show that the proposed framework overall improves activity recognition accuracy by up to 15 percent for the first dataset and by up to 32 percent for the second dataset.
Recent advances in mobile devices, wearable sensors and data analytics are enabling greater adoption of remote health monitoring systems, where human activity recognition (HAR) plays a key role. Most HAR systems assume that consecutive data segments are independent and aim at accurately classifying each individual data segment. However, because training and operating contexts are often different, and collecting labels can be time-consuming and expensive, the recognition accuracy is typically considerably lower in the new (operating) context. In this paper, we propose a framework for HAR that has several unique capabilities: 1) Turns predictions into useful information (e.g. daily distribution of activities for a user) that could be assessed by a clinician or health professional. Our goal is to mitigate the effects of variation between a training and an operating context. 2) Leverages the dependence between consecutive data samples to perform robust activity recognition. And 3) Performs activity classification for variable length set of activities. The framework we present encompasses two main modules. The first is a Recurrent Long Short-Term Memory deep neural network, the core component which gives us the capability to capture the dependence between consecutive sensor data samples and generate labels for every micro segment. The second is a post-processing step performed for uncertainty filtering and label smoothing of decisions. The filtering method enables the system to recognize activities in the presence of difference in the training and testing data distributions, and the smoothing method makes the system properly capture the distribution of activities. We demonstrate that our deep learning architecture can achieve 90% accuracy in standard training scenarios. Furthermore, we show that, in the case of inter-subject variation, the two methods allow us to still accurately capture the distribution of activities.
Mobile health monitoring plays a central role in the future of cyber physical systems (CPS) for healthcare applications. Such monitoring systems need to process user data accurately. Unlike in other human-centered CPS, in healthcare CPS, the user functions in multiple roles all at the same time: as an operator, an actuator, the physical environment and, most importantly, the target that needs to be monitored in the process. Therefore, mobile health CPS devices face highly dynamic settings generally, and accuracy of the machine learning models the devices employ may drop dramatically every time a change in setting happens. Novel learning architecture that specifically address challenges associated with dynamic environments are therefore needed. Using active learning and transfer learning as organizing principles, we propose a collaborative multiple-expert architecture and accompanying algorithms for the design of machine learning models that autonomously adapt to a new configuration, context, or user need. Specifically, our architecture and its constituent algorithms are designed to manage heterogeneous knowledge sources or experts with varying levels of confidence and type while minimizing adaptation cost. Additionally, our framework incorporates a mechanism for collaboration among experts to enrich their knowledge, which in turn decreases both cost and uncertainty of data labeling in future steps. We evaluate the efficacy of the architecture using two publicly available human activity datasets. We attain activity recognition accuracy of over (for the first dataset) and (for the second dataset) by labeling only of unlabeled data.