The Low Voltage (LV) distribution network (defined as 1kV and below) represents a significant proportion of network expenditure, yet until recently there has been relatively poor visibility of these assets.
This project will identify and test novel methodologies that could contribute to enhanced asset management for LV network assets by introducing predictive methods based on models that can determine the probability of failure. The quantitative analysis within the project will focus on LV cables, but we anticipate there will be broader learning for other asset types. A business case for implementing further development of the framework shall also be prepared.
Benefits
The output from this work will be the results from a prototype predictive framework that aims to predict the condition of the LV network and forecast the degradation of the network assets through-time based on a range of operating scenarios. Networks will have the opportunity to benefit from increased understanding of how LV assets, in particular, LV cables interact with their environment. In addition, initial results for the prediction of the current condition of the LV network and the forecasted degradation that may lead to short term improvement opportunities. A business case development for further research shall outline future development opportunities, with the eventual target of business as usual implementation.
Learnings
Outcomes
The full results for this work is detailed in the “LV Predict Final Report” which is available on the ENWL website.
Development of a modelling framework
Investigation of LV failure modes
Three areas of research were investigated to determine the most material LV network failure modes:
- Failure modes within the NaFIRs data - It is helpful to distinguish between two key types of failures: (i) those caused by the condition of the cables (primary drivers), and (ii) those caused by external influences (secondary drivers). The NaFIRS data suggests that most failures are due to primary drivers. Degradation-based failure mechanisms are the most common, with transient fault mechanisms (i.e. sporadic faults that result in a temporary loss of transmission) the second most common. Transient faults could be caused by several different mechanisms, such as water ingress or by tripping circuit breakers in the system. It is noted that this is a symptom of failure rather than a cause of failure. The third and fourth most common types of failure are third-party intervention and corrosion respectively.
- Learnings from in service experience - In-service experience of the LV network was provided by ENWL via a workshop session. It was established through anecdotal experience that approximately 95% of LV network failures occur at the joints, with relatively few failures occurring in the cables themselves. Further anecdotal evidence suggests that a significant number of failures occur after a change in weather (for example several days after rainy periods start or at the start of a warmer period).
- DFMEA - A failure tree was developed by the project team, incorporating subject matter expertise from ENWL, the existing knowledge of the Frazer-Nash and TNEI project team, and findings from a literature review. It shows, for example, that the condition of joints and cables is dependent on mechanical interference and environmental damage, as well as the initial condition of the joint/cable and history of previous failures.
The data gathered from the workshop, DFMEA, and NaFIRS analysis suggests that third-party intervention and thermal degradation of the joint are the most likely leading failure mechanisms for the underground LV network. Thermal degradation of the joint filler material can lead to water ingress, which could also lead to transient faults and damage to the conductors. In the modelling, prioritisation is given to the simulation of thermal degradation of the joint (and potential subsequent water ingress), as these failure modes are degradation-based and can be modelled as physical processes, unlike third-party failure causes, where the human factors are much harder to anticipate and predict. Hence, this could provide the most value as the key part of a predictive maintenance tool.
Constructing a database of all underground LV cables
The database was created by feeding data from ENWL’s GIS system into the “NetworkX Python” package to identify LV cable feeders, and the numbers of customers connected to each of them.
The basic approach has been to construct node and edge graphs of the LV cable system supplied by each HV and LV secondary substation. Customer addresses are not exactly connected to the LV services within the GIS data. Therefore, it is assumed that each customer address is connected at the end of the closest LV service line. Once these graphs have been assembled, it is simple to count how many customers are connected downstream of each cable. The GIS data also contains other information about the cables such as the type of conductor and its cross-sectional area.
The physical simulation and statistical modelling in this project have only considered the aggregated LV feeder, rather than consider every single LV cable segment. Therefore, the modelling is based on the design and cross-sectional area of the first cable segment connected to the secondary transformer and the total number of customers downstream of this segment, rather than all the individual segments. This will typically be the main pinch point that degrades most quickly due to thermal stress.
Outlining of modelling approach.
The probabilistic failure model developed within this project focusses on how demand causes cable temperature to vary, and how temperature cycles can then cause damage within the cable. Consideration is given to how estimates of demand could be improved by using a partial penetration adoption of smart meters, and by potentially using information about demand on higher voltage levels.
The basic approach to this modelling framework is:
- Map out the connection between the variables.
- Use physical models and limited available data to generate “training data” (e.g., for cable temperatures and damage).
- Fit univariate models to the training data for each variable, which describe the probability of observing different levels of demand, temperature, damage etc., with an emphasis on extreme values of those variables (e.g., 1-in-10-year levels of demand).
- Use data-driven probabilistic machine learning methods to combine these univariate models to a single multivariate model, while still accounting for the structure of the relationships between parameters.
Modelling customer demand and cable temperature
There are two main topics of interest when modelling customer demand and cable temperature:
- The extent to which probabilistic predictions can be made (that are possibly quite uncertain) about the nature of these critical extremes of active power (kW) demand for a group of customers served by a cable, based on the type of data about those customers that might typically be available to a DNO. Also, the best methodology to make these predictions.
- The extent to which these predictions about demand extremes can be taken and combined with other readily available data about a cable. For example, the type of soil in which it is buried, the altitude of its location etc – to produce (uncertain) probabilistic predictions about the extremes of temperature and rates of change in temperature that the cable is likely to experience. Again, the best methodology to make such predictions needs to be established.
Once these topics were explored and methodologies developed and verified, with the model was run using the database of ENWL underground LV cables.
To research these topics, suitable data was used; consisting of demand sequences (i.e., time series) for relatively small groups of domestic customers, for which basic data is available (i.e., annual energy consumption values). It is assumed that, in practice, the DNO would not have access to the former data (notwithstanding their ability to access smart meter data), but that a DNO would have access to annual consumption values.
By extracting key statistics from these time series, a statistical model can be fitted. How sequences of demand values for customer groups converts to a sequence of cable temperatures must be understood. Obtaining coincident time series of power demand and temperature enables the fitting of a statistical model capturing the relationship between extreme demands and extreme changes in demand, and also extreme temperatures and extreme changes in temperature. While such a model is entirely statistical, acquiring the coincident time series of temperature, given a demand time series, requires physical modelling – specifically, a physics-based cable temperature model. The same need for complementary physical and statistical models is true as the model development progresses from temperature to damage and from damage to risk of imminent failure.
Physical Modelling of LV network degradation
Physical joint thermal stress modelling
Heating of LV cables and joints can lead to differential thermal expansion due to the interface between different components, which can result in thermal stress applied to parts of the LV network. To understand this effect in more detail, results from the cable and joint temperature model are fed into a joint stress model. Specific interest is taken in the cable joints, as operational experience suggests that this is where 95% of failures occur in the network.
Physical joint damage modelling
Stresses in the joint can damage the joint material, leading to eventual joint failure. To understand joint failure in more detail, stresses from the joint stress model are first converted from a continuous time series into discrete stress cycles Next, the discrete stress cycles are used as part of a physical model of joint damage. Finally, a sensitivity analysis of temperature on the physical joint damage model is carried out.
Physical modelling of other environmental factors
The probabilistic modelling framework for the LV network also considers other environmental effects. Work in this project has focused on understanding the influence of electrical-based degradation and chemical-based degradation on the LV network.
Calculating remaining life using the physical model
Once the physical damage has been calculated using the physical damage model, the damage results can be used to calculate the remaining life in the underground LV assets. Future life calculations require knowledge of the asset age, previous loading history, expected future loading, and expected degradation rate. Currently, some of this information is unknown, therefore assumptions must be made when calculating the remaining asset life. Several methods are proposed here to calculate the remaining asset life, each of which have different benefits and drawbacks.
Overall sensitivity analysis of the physical damage model
A global sensitivity analysis is performed to understand how each variable for the underground LV asset may influence the asset damage.
Statistical modelling of LV Network degradation
Statistical cable damage modelling
The final model in the statistical modelling chain linking demand, temperatures, and damage converts high quantiles of temperature to a predicted probability distribution for damage sustained within a year. Target values of annual damage were calculated using the physical models and capped at one, where a damage of one represents a failed cable. For simplicity, only total damage has been considered, but it may be more appropriate to jointly predict a bivariate distribution of plastic damage, and fatigue plus creep damage, due to the differing ways in which these forms of damage are sustained.
Statistical life prediction modelling
From the model for damage, it is possible to determine the probabilities of different levels of a cable’s remaining useful life. Of significance is the ability to calculate the probability of useful life reducing below a certain threshold over a specified duration (e.g., less than a year of useful life within the next five-year price control). In addition, it has been shown that a partial adoption of smart meters can reduce the uncertainty associated with all predictions (demand, temperature and damage). This functionality emerges from the ability of the model to predict annual damage sustained by a cable with a specific number of customers, of a specified construction, with soil thermal resistivity estimated from the cable location. By simulating the damage for a sequence of years, perhaps under a scenario for growing LCT adoption or changing annual energy consumption, a large set of samples of a sequence of damage can be produced.
Lessons Learnt
A number of high level observations were highlighted during the project:
- Different soil types have a strong influence on the cable life. The cables are particularly sensitive to peat, as it has a low thermal conductivity. This means temperature cannot easily dissipate from the cable when it heats up, thus causing damage and reducing the remaining life. Cables aren’t often buried in peat, but in some geographical conditions it may be the only real option for network planning.
- Cables are sensitive to changes in ground temperature. Increased soil temperature increases the cable temperature, which increases cable damage and leads to a reduction in cable life.
- Increasing the proportion of customers that have EVs can, without smart charging, significantly increase cable damage and lead to a significant reduction in cable life. In this case, 10% EV usage can provide a life of over 100 years, but when the EV proportion is raised to 100% then the life will be reduced to approximately a decade for a standard soil composition.
- In most cases creep damage is the dominant failure mode for the cable joints.
- Fatigue damage is extremely sensitive to the initial crack size in the cable joint. This indicates that under certain conditions, the presence of defects in the joint filler material could influence the degradation of the cable joint.