Distribution network monitoring has the potential to improve service levels by reporting the origin of fault events and informing the nature of remedial action. To achieve this practically, intelligent systems to automatically recognize the cause of network faults could provide a data driven solution, however, these usually require a large amount of examples to learn from, making their implementation burdensome. Furthermore, the choice of input to such a system in order to make accurate classifications is not always clear. In response to this challenge, this paper contributes a means of using minimal amounts of historical fault data to infer fault cause from substation current data through a novel structural similarity metric applied to the associated power quality waveform. This approach is demonstrated along with disturbance context similarity assessment on an industrially relevant benchmark data set where it is shown to provide an improvement in classification accuracy over comparable techniques.

Increased observability on power distribution networks can reveal signs of incipient faults which can develop into costly and unexpected plant failures. While low-cost sensing and communications infrastructure is facilitating this, it is also highlighting the complex nature of fault signals, a challenge which entails precisely extracting anomalous regions from continuous data streams before classifying the underlying fault signature. Doing this incorrectly will result in capture of uninformative data. Extraction processes can be confounded by operational noise on the network including harmonics produced by embedded generation. In this paper, an online model is proposed. Our Bayesian Changepoint Power Quality anomaly Segmentation allows automated segmentation of anomalies from continuous current waveforms, irrespective of noise. Demonstration of the effectiveness of the proposed technique is carried out with operational field data as well as a challenging simulated network, highlighting the ability to accommodate noise from typical network penetration levels of power electronic devices.

Cables are the largest assets by volume on power distribution networks and the assets with the least health information routinely gathered. Projections over the next 8 years suggest increased penetration of low-carbon technology (LCT) at the distribution level with higher loads resulting from electric vehicle (EV) and heat pump uptake. Over this period, increased cable loading will directly influence their lifetimes and may mean that existing asset management practices need to be revised to understand the specific impact on end-of-life assessment. Accordingly, this paper uses a physics-based thermal lifetime model based on projected uptake trends for LCTs to evaluate the impact on distribution cable lifetime. Two case studies are presented considering portions of network and the ultimate impact on asset life over the next decade. Two commonly used cables are considered to quantify the lifetime reduction caused by LCT for asset fleets. The paper shows that the projected uptake of EVs and heat pumps will have a detrimental effect on cable life with a 30% reduction in cable lifetime possible.

This paper proposes the use of Gaussian Process Regression to automatically identify relevant predictor variables in a formulation of a remaining useful life model for unmonitored, low value power network assets. Reclosers are used as a proxy for evaluating the efficacy of this method. Distribution network reclosers are typically high-volume assets without on-line monitoring, leading to an insufficient understanding of which factors drive their failures. The ubiquity of reclosers, and their lack of monitoring, prevents the tracking of their individual remaining life, and, confirms their use in validating the proposed process. As an alternative to monitoring, periodic inspection data is used to evaluate asset risk level, which is then used in a predictive model of remaining useful life. Inspection data is often variable in quality with a number of features missing from records. Accordingly, missing inputs are imputed by the proposed process using samples drawn from an advanced form of joint distribution learned from test records and reduced to its conditional form. This work is validated on operational data provided by a regional distribution network operator, but conceptually is applicable to unmonitored fleets of assets of any power network.

Fault analysis based on high-resolution data acquisition is growing in use as it offers a more complete picture of faults which provides an opportunity to deal with failures more effectively. However, with increased volume of data collected, it becomes impossible for engineers to interpret every fault instance. A machine learning approach to classification should be the solution to this, but it is time-consuming to manually label faults for training and validation making data-driven approaches impossible to transfer into practical implementation. A solution to this is to unify fault analysis with maintenance report analysis to automate the generation of training labels. This paper outlines how a fully automatic fault detection and diagnostic approach based around power quality waveform analysis can be used to improve situational awareness on distribution networks. The methodology is illustrated using operational case study data and realistic simulations to demonstrate the diagnostic functionality as well as the practical benefit. In particular, classification accuracy is shown to approach that of expert labelled fault data.