Virtually all power distribution circuitry in the United States operate in a multi-earth-grounded configuration. This means that an electrical fault will exist if an electrical connection is made to ground. If the normal growth of vegetation under or next to distribution lines is not held in check by pruning, growth will eventually bring it close enough to the wires to create a conductive path to earth; causing a fault which will lead to an outage.

To ensure high levels of reliability in the distribution system, vegetative maintenance (tree trimming and other vegetation control measures) is periodically conducted by electric utilities. Since the costs of such maintenance counts a large fraction of the total amount spent on distribution system maintenance [1] and can be millions of dollars even for a small utility [2], the cycle length is generally selected by economics. It might seem like a longer trimming cycle would be the most economical choice because it spreads the cost out over time. However, this thinking is flawed because the amount of work that has to be done increases very rapidly with time. The work load increases due to two factors. First, the amount of biomass that must be cut and handled increases rapidly with each year of growth and second, as the tree grows closer to the energized lines, workers must use caution to keep from getting hurt or killed. In fact, short cycles are expensive due to the amount of work and long cycles are also costly due to excessive biomass and loss of productivity. Therefore, an advanced understanding on the relationship between vegetative growth and system reliability can help the power companies determine an effective vegetation maintenance schedule.

Some studies have been made to quantitatively analyze the effects of vegetative growth on distribution system outages. The paper [3] presented a time series and a non-linear machine learning regression model for predicting the number of vegetation-related outages that occur in power distribution systems on a monthly basis. In the article [1], several direct failure-rate models were investigated to predict the time-varying, vegetation-related failure rates of overhead distribution power lines. The authors in [4] developed statistical models for estimating the impacts of tree trimming on electric power system outages under normal operating conditions. Another study [5] reported a maintenance-scheduling algorithm that determines the optimal location and time for performing vegetation maintenance on overhead distribution feeders using a vegetation failure rate model. However, the number of literature on this critical issue is still limited due to the difficulty of collecting outage data, therefore this problem has not been investigated to the fullest extent yet.

In this paper, a novel approach to quantify the impact of vegetation growth on electric power system outages is proposed. Utilizing statistical and machine learning predictive models, the proposed method advances the understanding on the relationship between vegetative growth and system reliability and enables effective and timely decision-making actions by power industry. What distinguishes this approach from the other studies in the literature could be summarized in the following three aspects:

• From a theoretical perspective, to the best of our knowledge, ours is the first study that includes some statistical models such as zero-inflated Poisson and zero-inflated Negative Binomials, as well as the mixed models into the analysis of actual outage data. Considering a majority of 0’s in the data and the clustering nature, those models provide better fit to the occurrence of outages than the existing methods.

• One main reason for the perceived lack of studies on this critical problem is the limited access to a sufficient amount of outage data. Our approach takes advantage of a considerable amount of vegetation-related outage data that is provided by a power company in Midwestern United States, thus allows a sound statistical basis and strong conclusions to be drawn.

• In practice, our study explicitly suggests an optimal cycle length of every 6 years, which could be greatly helpful for power companies in devising a cost-effective schedule, improving system reliability, and maintaining customer satisfaction.

The rest of this paper is organized as follows. Section 2 describes the outage data. In Section 3, generalized statistical linear regression models, including Poisson, Negative Binomial, Zero-Inflated, and their mixed model variants are introduced and developed. In Section 4, we fit the statistical models on a real outage data, the results are presented and discussed. The paper ends with discussion and future research directions.

The analysis in this paper is based on data from a power company in Midwestern U.S, containing 431 vegetation-caused outages on 144 circuits for 2012 through 2016. A vegetation-caused outage is defined as any outage caused when the vegetation gets close enough to sway into the line or to create a path for a tree dwelling animal to bridge the air-gap between energized lines and vegetation, under normal (storm exclusion) operating conditions. For this data, the date and time of each outage, the duration of the outage, the name of the circuit where the outage occurred, the number of years since the last routine vegetation maintenance was performed (year of vegetative growth), and the number of customers affected by the outage are recorded. Figure 1 shows the histogram of vegetation-caused outages on each circuit. It can be seen that most circuits only have 1 outage, and the distribution is right skewed.

In statistics, count data is a type of data in which the observations can take only the non-negative integer values, for example, the number of power outages on a circuit in this paper. When modeling count data, the classical ordinary least-squares (OLS) regression is often inappropriate because the homoscedasticity and normality assumptions are violated. The violation of the basic OLS assumptions can result in inaccurate estimates of standard errors, and misleading p-values and consequent confidence intervals [6]. A class of generalized linear regression models has been developed for modeling count data. These models have a number of advantages over an ordinary linear regression model, including a skew, discrete distribution, and the restriction of predicted values to non-negative numbers [7].

The most widely used model for count data is Poisson model. The Poisson model is made up of a Poisson probability mass function (PMF) denoted as P (y _i = k) used to calculate the probability of observing k events given a mean event rate of λ, and a link function that is used to express the mean rate λ as a function of the regression variables X . The Poisson PMF is P ( y i = k ) = e − λ i λ i k k ! , i = 1,2 , … , n , (3.1) where y _i is the observed count for the ith row in the dataset, λ _i is the event rate corresponding to the ith sample. The link function of the Poisson regression model is expressed as ln ( λ i ) = x i β , (3.2) where  ln (⋅) is the natural logarithm function, x _i = [x _i1, x _i2, … , x _ip ] is the regression variables in the ith row, and β is the vector of regression coefficients.

The Poisson model assumes that the mean and variance of the errors are equal, but usually in practice the variance of the errors is larger than the mean. In these cases, we say that the data is “overdispersed”. When overdispersion arises, the Poisson model is not proper, and an alternative is a Negative Binomial (NB) model. The Negative Binomial distribution is a form of the Poisson distribution in which the distribution’s parameter is itself considered a random variable. The variation of this parameter can account for a variance of the data that is higher than the mean. With a NB model, the count data are assumed to follow a NB PMF as what follows, P ( y i = k ) = Γ ( k + 1 / α ) Γ ( k + 1 ) Γ ( 1 / α ) 1 1 + α λ i 1 / α α λ i 1 + α λ i k , (3.3) where Γ(⋅) is the gamma function, α is the overdispersion parameter, and the regression model is same as for the Poisson model in (3.2).

When count data has both excess zeros and large counts, zero-inflated Poisson regression (ZIP [8]) is a practical way to deal with such situation. It assumes that with probability p the only possible observation is zero, and with probability 1 − p a Poisson (λ) random variable is observed. The intuition behind the ZIP model is that there is a second underlying process that is determining whether a count is zero or non-zero. A ZIP distribution can be written as P ( y i = k ) = p i + ( 1 − p i ) e − λ i if k = 0 ; ( 1 − p i ) e − λ i λ i k k ! if k = 1,2 , … (3.4) The Poisson mean λ _i and the probability p _i are linked to the explanatory variables through the log link in (3.2) and logit link as l o g i t ( p i ) = z i γ , (3.5) where z _i is the vector of covariates for the ith subject, and γ is the vector of the corresponding regression coefficients.

The zero-inflated Negative Binomial (ZINB) regression is used for count data that exhibits overdispersion and excess zeros. The data distribution combines the NB distribution and the logit distribution. The model can be expressed as P ( y i = k ) = p i + ( 1 − p i ) g ( y i = 0 ) if k = 0 ; ( 1 − p i ) g ( y i ) if k = 1,2 , … , (3.6) where g (y _i ) = P (y _i = k) is defined in (3.3), and the link functions are the same with the ZIP model in (3.2) and (3.5).

In this section, we fit the models on the outage data set and draw conclusions on the basis of these fitted models. The Akaike Information Criterion (AIC [9]) and Bayesian Information Criterion (BIC [10]) are utilized to compare models. The AIC of a fitted model is defined as A I C = − 2 ∗ ln L ̂ + 2 ∗ k , (4.1) and the BIC is expressed as B I C = − 2 ∗ ln L ̂ + log ( n ) ∗ k , (4.2) where L ̂ is the maximum value of the likelihood function for the model, k is the number of estimated parameters in the model, n is the sample size. In comparing models, a smaller AIC and BIC is favorable.

This paper investigates the role of vegetative growth on distribution reliability. Some statistical models including Poisson, Negative Binomial, Zero-Inflated models and their variants are utilized. Based on this study, the following conclusions can be drawn.

1) There is a significant statistical relationship between the reliability of the electric power distribution system and the number of years of vegetative growth on distribution circuits.

The study can advance the understanding on the relationship between vegetative growth and system reliability and should be useful to help the power companies determine an effective vegetation maintenance schedule. However, there are three main limitations of the models and future research could be conducted to obtain more accurate results.

First, we limited the attention on the effect of vegetative growth on power system reliability, other possible factors are not included due to lack of access to related data. As a result, the performance of the analysis may be improved by the inclusion of additional climate and geographical information. The second limitation lies on the fact that the models are fit with data from only normal (storm exclusion) operating conditions, which means that our models might underestimate the benefits of tree pruning. Additional data and study would be needed to learn the impacts of tree maintenance on distribution system reliability under storm conditions. The third limitation is that the analysis is based on data from only one company, consequently the model may not fit the data from another power company, especially if the region where the tree types, temperature, population density, wind regimes, and precipitation patterns is substantially different from Midwestern United States. If more data are available, it will further test the effectiveness of the statistical models, and enhance insights on the importance of vegetative maintenance in improving power system reliability.