A Technique for Transformer Remnant Cellulose Life Cycle Prediction Using Adaptive Neuro-Fuzzy Inference System

Abstract

This article presents an ultramodern modelling algorithm for predicting the remnant cellulose life cycle for oil-submerged power transformers based on the adaptive neuro-fuzzy interference system (ANFIS). The polymer characteristics, degree of polymerization (DP), and 2-furaldehyde (2FAL) of 100 power transformers were measured and collated, which were apportioned into 70 training databanks and 30 as testing datasets. The remnant cellulose life cycle of the transformer was predicted using the proposed ANFIS model characterized by polymer characteristics, DP and 2FAL as inputs. The proposed approach returns 98.23% training and 99.86% testing reliability. The proposed model was applied to 10 transformer case studies in predicting their remnant cellulose life cycle. To corroborate the proposed ANFIS, a comparative study was carried out by employing existing approaches in predicting the remnant life cycle of the case studies, and significant error margins were observed. At large, the results presented in this article certify the dominance of the proposed ANFIS algorithm over compared models. The proposed ANFIS furnishes a pathway to obliterate the constraints of classical techniques in evaluating the transformer DP and remnant cellulose life cycle.

1. Introduction

The health index of electrical power transformers that are in service inherently depreciates on account of the decomposition process of insulating materials. During the transformer in-service lifespan, miscellaneous loading conditions can accelerate the decomposition process of cellulose and oil-insulating materials by initializing a surge in ageing agents viz. oxygen, temperature and moisture [1]. These agents will induce the inception of acids, which may adversely affect the health index of the cellulose insulation. The latter may trigger electrical transformers to not achieve the intended in-service lifespan. The knowledge of the remaining lifespan is thereby a compelling characteristic for utility owners. As a result, the ageing of transformers has become a major concern to utility owners. The insulating materials in transformers will incessantly decompose throughout life, and notably, the cellulose insulation cannot be substituted. The operational boundaries of the electrical transformers are constrained by the decomposition of cellulose insulation which decrements the mechanical strength of the insulating cellulose [2].

Staunch and efficacious condition monitoring and indicative approaches must be exercised to steer clear of impromptu outages nationwide. Incipient identification and unearthing of dissolved gases in the insulating transformer oil have acclimated to the fastest-growing practice in the identification of incipient transformer faults [3,4,5].

The dissolved gas analysis (DGA) uses a variety of gases that are generated based on oil putrefaction and the DP of the cellulose insulation [6,7]. Nevertheless, it is infeasible to acquire cellulose paper samples from in-service transformers, particularly from hotspot regions. Subsequently, numerous interrelations between the DP and furanic compounds released into the insulating oil because of the decomposition of the cellulose insulation have been reported in the literature [8,9]. These interrelations have so far proven to be inconsistent as they contribute to distinct values for similar furan concentrations.

The contribution of current research: In this article, an ANFIS modelling algorithm for predicting the cellulose insulation remnant life cycle in electrical transformers was developed to outmanoeuvre the constraints of existing DP and loss-of-life models. Polymer characteristics, consisting of 100 DP and 2FAL databanks, were collated to develop and corroborate the performance of the proposed ANFIS model against the actual data. A comparative assessment of both the proposed ANFIS and existing models was carried out for 10 transformer case studies. It has become evident that the proposed ANFIS model is superlative when tested using the databank and against existing models.

The novelty of current research: The main aim of this work is to predict the cellulose insulation remnant life cycle in transformers. Notwithstanding that numerous researchers have worked on the transformer remnant cellulose life cycle prediction predominantly using in-oil dissolved gases (hydrogen, methane, ethane, and acetylene), miniature research works [8,9,10] have been reported about the prediction of remnant transformer life by unifying the parameters, DP and 2FAL. The rate of degree polymerization of the cellulose insulation is the principal criterion for designing the prediction approach for the remnant cellulose lifetime. Three DP models that are widely adopted in the literature are considered and ascertained to yield a significant error margin when compared to each other and against the practical dataset values. These data are particularly valuable in the development of a reliable prediction approach for the cellulose insulation remnant life cycle of transformers. Numerous investigators compared the existing DP models in their work. Even though there are comparable works, nevertheless, in the present study, remnant cellulose life of the transformer is achieved from an uncommon ANFIS modelling algorithm and its performance is studied. There is currently no existing literature that covers the prediction of remnant life using 2FAL and DP. ANFIS offers significantly better learning capability: for the same network intricacy, a significantly lower mean absolute deviation (MAD), root mean square error (RMSE) and mean absolute percentage error (MAPE) compared to existing models is attained.

The manuscript structure: This manuscript is organized as follows: Section 1 offers an introduction to transformer cellulose life, contribution and novelty. Section 2 provides the applicable works in relation to the prediction of transformer cellulose life using artificial intelligence techniques. Then, Section 3 covers the fundamental principle of the existing DP models, the proposed approach and some statistical indicators. After, Section 4 presents the results of the proposed approach and examines its performance against existing models, then also presents various case scenarios by adopting existing methods and the proposed approach. Section 5 provides a discussion of the results, and the manuscript is concluded in Section 6.

2. Applicable Works

Computational techniques have been proposed by several researchers to circumvent the constraints introduced by existing empirical models. Notwithstanding the account that furan is the utmost attainable yet dependable transformer cellulose paper valuation, this measurement is not carried out sporadically by the power utilities. To ascertain the up-to-date state of paper insulation, additional inexpensive methods are required. A machine learning algorithm can be utilized to model the up-to-date transformer cellulose paper status. Numerous studies have been carried out to investigate the potential of this technique.

The assessment of transformer paper condition has been studied using fuzzy logic and ANFIS models [10,11], support vector machine (SVM) [12], artificial neural network (ANN) [13] and regression-based models [14]. These techniques have respective merits but with some constraints such as over-sampling and data uncertainty, which can be addressed by swarm optimization techniques [15] and probabilistic methods [16,17].

In [18], fuzzy logic was employed to predict the transformer remnant lifespan. In [19], the k-nearest neighbours algorithm (k-NN) and decision trees were applied in predicting the levels of furan concentration of oil-immersed transformers. The transformer degree of polymerization was predicted using ANFIS and, additionally, the projected life valuation [20], and a multiple regression model was also developed as a benchmark against the ANFIS model [21]. SVM is among the regularly implemented machine learning algorithms for dataset categorization [21,22]. SVM is utilized in [23] to predict electrical load in conjunction with other algorithms such as fuzzy time series (FTS) and harmony search (HS). In transformer diagnostics, SVM was applied [24,25] for identifying faults. Numerous machine learning methods were employed in [26]; here, SVM was developed together with ANN, k-NN, decision trees and the naive Bayes classifier to evaluate the furan concentration in transformers. These works asseverate with comparatively low precision on the SVM classifier. Currently, there is no scientific approach existing to determine the state of the end-of-life of a transformer that is in service. The composite of analytical methods, routine inspection and test approaches, when applied collectively, support forming a comprehensive picture of the state of units in the field.

This work contrivances prediction analysis by means of ANFIS as a supplementary insight to support power utilities and transformer manufacturers evaluating transformer cellulose paper insulation condition by means of dielectric oil measurements data, i.e., 2FAL. The foremost concerns of developing ANFIS model prediction analysis are reviewed, those being data preparation and model authentication to unearth the preeminent model fashioned to fulfil the veracity level anticipated. The proposed model is then benchmarked to conventional DP and loss of life (LOL) models to corroborate the prediction result.

3. Materials and Methods

3.1. Existing Models

It is broadly recognized that the health status of a transformer is heavily reliant on the holistic condition of its insulation system. The decomposition of the oil-submerged transformer insulating system divisible into cellulose paper and liquid insulating oil eventuates through a convoluted multiplicity of factorial activities traceable to the interrelations between ageing agents and their cumulative effects. The additory impact of the rise in temperature, dispersion of moisture, generation of acids and formation of oxygen bubbles within a transformer is characterized as the underlying root of the transformer ageing mechanism. Painstaking investigations were carried out to institute the interrelationship between 2FAL and DP [27,28], the latter in conjunction with their corresponding cellulose remnant life cycle models underscored by some of the proposed models in literature as shown in Equations (1)–(4).

The DP test is conducted to measure accurately the decomposition of the cellulose paper insulation in the transformer. Cellulose, which is employed as a solid insulating material, is a one-dimensional polymer fragment characterized by masses of glucose units. The cellulose strings are accentuated by the average number of DP glucose molecules. The value of the DP declines over time correlative to the fragmented cellulose molecule. The degradation rate is wholly reliant on the thermal condition. It is generally acknowledged that the DP values change from approximately 1200 for new cellulose paper to no less than 100 at the end of life [29].

At a DP of approximately 200, the cellulose paper has succumbed to about 70% of the original durability. At this juncture, the cellulose paper becomes delicate, and the unit in-service is categorized to have reached the end of its operational life. The level of furan contents is a particularly significant ageing measure for the estimation and life cycle valuation of units in service [30]. Furanic compounds are simpler to ascertain as compared to DP and TS (tensile strength) for the reason that only testing of the oil sample is required than extracting physical paper samples of the unit in-service.

Furan compounds are derivatives of the ageing mechanics of cellulose insulating material during transformer service. It has been well recognized specifically that 2FAL has a strong correlation to the DP. Prevalent existing models are expressed in Equations (1)–(4). This model was proposed by Chendong et al. [31] and expressed as follows in Equation (1):

$$\mathrm{DP}=-285.714\times \left({\log }_{10}\left(2\mathrm{FAL}\right)-1.51\right)$$

(1)

This model was proposed by De Pablo et al. [32] and is expressed as follows in Equation (2):

$$\mathrm{DP}=\frac{1850}{2\mathrm{FAL}+2.3}$$

(2)

This model by Vaurchex et al. was reported in [33] and is expressed as follows in Equation (3):

$$\mathrm{DP}=\frac{2.6-{\log }_{10}\left(2\mathrm{FAL}\right)}{0.0049}$$

(3)

The loss of life model expressed as a function of the DP has been proposed in [34] and is expressed as follows in Equation (4):

$$\mathrm{LOL}=\frac{41}{2}\times \ln \left(\frac{1100}{\mathrm{DP}}\right)$$

(4)

3.2. Proposed ANFIS Modelling for Cellulose Remnant Life Cycle

An adaptive neuro-fuzzy inference system (ANFIS) is a class of ANN founded on the Takagi–Sugeno–Kang fuzzy inference [35]. Considering that this method incorporates NNs and fuzzy logic (FL) fundamentals, it has the capability to possess the strong point of respective fundamentals in one architecture. The inference system is consonant with a class of fuzzy computerized IF–THEN rules which have the ability to learn and estimate nonlinear functions. Consequently, ANFIS is deemed to be a general estimator.

The proposed ANFIS model for predicting the remnant cellulose life cycle of a transformer is illustrated in Figure 1. The following steps were carried out to develop the mode:

3.2.1. Data Preparation

In the first step, the collated data from 100 transformer samples were partitioned into training and testing datasets: 70 datasets were used for training, and 30 datasets were used to test the proposed model.

3.2.2. Proposed ANFIS Modelling

The training databank of 100 transformer samples is used to develop the proposed fuzzy logic ANFIS model, as shown in Figure 1. The subsequent step is to unearth the optimum grouping of the membership function (MF). The total and category of MF for respective input variables are designated, the model of the respective combination is trained to utilize the 70-training databank and, additionally, the respective model is tested to ascertain the suitable combination.

3.2.3. Accuracy Evaluation

The proposed model is then tested using the 30-testing datasets, and thenceforth, the error is computed for the overall databank.

The relationship between the 2FAL and DP is demonstrated in Figure 2. It can be observed that the curve reaches a steady state at about 250 DP. The latter might indicate the end of life for the cellulose paper insulation.

The latter can be substantiated by Figure 3. Here, the consumed life is compared with the amount of DP that is present in the paper. It can be observed that at a DP of about 250, the consumed life is about 30 years, a standard lifetime for a transformer in service.

Numerous measures can be employed to appraise the performance of the proposed model against actual data and for the existing methods against actual data. In this work, the veracity of the proposed model has been appraised using the statistical indicators indicated below:

3.3. Mean Absolute Deviation (MAD)

This error evaluation criterion measures the difference between the observed and predicted LOL. Equation (5) evaluates MAD, where ${\mathrm{OV}}_{\mathrm{i}}$ and ${\mathrm{PV}}_{\mathrm{i}}$ are observed and predicted values, respectively [36]:

$$\mathrm{MAD}=\frac{1}{\mathrm{N}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}\left|{\mathrm{OV}}_{\mathrm{i}}-{\mathrm{PV}}_{\mathrm{i}}\right|$$

(5)

This is the average difference of the observed and predicted LOL values of the $\mathrm{i}$ unit, where $\mathrm{N}$ is the size of the column vector of the databank used to develop the proposed model.

3.4. Mean Absolute Percentage Error (MAPE)

This benchmark appraises the forecast error of the observed and predicted LOL values by employing Equation (6) [36]:

$$\mathrm{MAPE}=\frac{1}{\mathrm{N}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}\frac{\left|{\mathrm{OV}}_{\mathrm{i}}-{\mathrm{PV}}_{\mathrm{i}}\right|}{{\mathrm{OV}}_{\mathrm{i}}}$$

(6)

3.5. Root Mean Squared Error (RMSE)

The RMSE is a measure of veracity to compare the predicting errors of different models for a specific variable and not between variables since it is scale-reliant. This is a measure that minimalizes the variation of the inaccuracy distribution, and it is computed as follows in Equation (7) [36]:

$$\mathrm{RMSE}=\sqrt{\frac{1}{\mathrm{N}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{PV}}_{\mathrm{i}}-{\mathrm{OV}}_{\mathrm{i}}\right)}^2}$$

(7)

4. Results

4.1. Transformer Life Cycle Model Results

During the years of the transformer’s operational lifetime, various ageing indexes act interactively, making the ageing procedure intricate. As a result of this intricacy, developing a comprehensive mathematical model for insulation dilapidation mechanics is inconceivable. In this section, ANFIS is applied to predict transformer consumed lifetime and also to furnish an enhanced asset managing assessment tool, which produces high veracity in the model output prediction. The archetypal construction of the fuzzy inference system is expounded in [28]. The fuzzy inference system is generally adapted to model systems whose rule structure is identified based on the insight from characteristics of the presented data. The parameters of membership functions are elected arbitrarily by only observing the dataset. With the application of the ANFIS technique, the membership functions to the values of input and output datasets can be customized to rationalize wholly their features and adaptations. Membership functions in a fuzzy logic model are outlined by the parameters controlling the form of the respective membership function and the period covered. Conversely, the ANFIS technique facilitates dynamic fine-tuning of the membership functions by means of amending their parameters with every alteration in the input or output data datasets. Consequently, in the instance of condition monitoring assessment, membership functions and fuzzy guidelines could be adapted based on the disparities in the input variables and response from output data.

In the proposed model in this work, input data incorporating 2FAL content in the oil and the DP in the cellulose paper were collected from various dielectric oil submerged transformers of different voltage systems and power ratings, operational settings and lifetimes are considered. The LOL of the unit during service is reliant on the 2FAL content present in the oil and the corresponding DP thereof. Thus, these parameters were employed in developing a proposed model with the postulation that the 2FAL and DP parameters exist.

Gathered databank samples were apportioned into two clusters for training and testing, encompassing 70 and 40 comprehensive measures, respectively. The ANFIS technique is used to fine-tune the membership function parameters by means of the Levenberg–Marquardt backpropagation optimizing procedure. Employing empirical data of input and output variables, the set of rules to be followed in the computation of the problem-solving operations augments the parameters of membership functions by using weights that are calculated and modified by way of the learning procedure to curtail the error margin between the real and predicted results. Figure 4 illustrates the training error in years, which is the difference between the actual and predicted loss of life (LOL) of the units throughout databank training. It can be observed in Figure 4 that the training error declines significantly to a value of less than 0.005 at the iterance step (epoch) of about 280.

The framework of the proposed ANFIS model is shown in Figure 5 in which input variables to the system are 2FAL concentration in the dielectric oil and DP within the cellulose paper, while the output variable is the transformer predicted LOL in years.

As demonstrated in the proposed model structure above, the input variables (2FAL and DP) are charted by means of input membership functions and successively by way of output membership function into output variable (LOL). After the databank training ends, the product will be the developed ANFIS model and its augmented membership functions. The membership functions of the proposed ANFIS model are presented in Figure 6 and Figure 7. Conversely to fuzzy logic modelling, these membership functions are augmented each time ANFIS training is implemented, which enables incessant enhancement in the proposed model accuracy.

A mathematical illustration of the computerized rules related to the proposed ANFIS model is expressed as follows:

• if (input1 is in1mf1) and (input2 is in2mf1) then (output is out1mf1) (1)
• if (input1 is in1mf1) and (input2 is in2mf2) then (output is out1mf2) (1)
• if (input1 is in1mf1) and (input2 is in2mf3) then (output is out1mf3) (1)
• if (input1 is in1mf2) and (input2 is in2mf1) then (output is out1mf4) (1)
• if (input1 is in1mf2) and (input2 is in2mf2) then (output is out1mf5) (1)
• if (input1 is in1mf2) and (input2 is in2mf3) then (output is out1mf6) (1)
• if (input1 is in1mf2) and (input2 is in2mf1) then (output is out1mf7) (1)
• if (input1 is in1mf3) and (input2 is in2mf2) then (output is out1mf8) (1)
• if (input1 is in1mf3) and (input2 is in2mf3) then (output is out1mf9) (1)

Here, inputs 1 and 2 are the 2FAL and DP, while output1 is the predicted LOL. Through the ANFIS process, the parameters describing the shape of membership functions are evaluated by the mathematical operation repeated over the progression of the respective epoch and in reference to patterns prevailing in the introduced databank. This leads to a more veracious estimate of the practical data, which is a benefit over fuzzy logic modelling. The overall range of respective parameters employed in the proposed ANFIS model is elected based on confounded by respective membership functions and is evaluated by the augmented parameters computed by the ANFIS training technique. Figure 8 illustrates a graphical representation of the computerized rules related to the established ANFIS model.

The surface view of the proposed model is demonstrated in Figure 9.

To evaluate the veracity of the proposed model, the testing dataset was employed to benchmark the transformer consumed life cycle predicted by the proposed ANFIS model against existing models for predicting the DP. The performance in predicting the consumed life of the units attained employing the proposed model, and that projected by the existing methods, has been tabulated in

Table 1. Data Bank Statistical indicator.

Metric	Year	MAD	RMSE	MAPE
ANFIS Model	2023	0.019	0.152	0.091
Chendong [31]	1991	1.371	1.207	6.958
Vaurchex [32]	2003	5.204	2.424	21.142
De Pablo [33]	2007	2.609	1.667	13.101

.

By the above benchmarking, it can be conclusively substantiated that the proposed model has a superlative capacity to predict the transformer consumed life with the highest veracity.

This high degree of accuracy is described in the enhancement of membership function parameters when using ANFIS. These parameters are subjectively chosen when the FIS model is developed.

4.2. Case Studies

In this subsection, a comprehensive dataset composed of 10 oil samples (particularly of the 2FAL and DP values) of refurbished transformers is presented. The voltage levels range from 33 kV to 132 kV, and the power range is 15 MVA to 120 MVA. The units were located in various sites within South Africa. It must be emphasized that the dataset has not participated in the databank employed in the previous section to develop the proposed model. Some sets of data may be related to a transformer in different time intervals. Diagnostic tests linked to furan compound analysis and degree of polymerization are conducted and tabulated in

Table 2. Transformer case studies.

Transformer	2FAL	DP
1	1.72	380
2	2.021	360
3	2.374	340
4	2.789	320
5	3.277	300
6	3.851	280
7	4.524	260
8	5.315	240
9	6.245	220
10	7.337	200

. The transformers of the dataset are oil submerged with mineral oil.

Consequently, applying such a thorough dataset devises the results of the model taken into consideration as realistic. In the subsequent results, the model closest to the prediction of the actual LOL can be conclusively signified as a reliable model trained from the thorough databank optimally and predict LOL unerringly. In this work, two illustrious parameters from the furan test and DP test of transformer dielectric oil–cellulose paper insulation, including 2FAL and DP, are delineated as the input parameters of the proposed model. The predicted LOL parameter is the output of the proposed model.

A comparative analysis of the predicted LOL using the predicted method and the existing methods is tabulated in

Table 3. Predicted Loss of Life (in years).

Unit	Actual	ANFIS Model	Chendong [31]	Vaurchex [32]	De Pablo [33]
1	21.06	21.69	22.66	16.89	17.86
2	22.18	22.79	23.82	17.51	19.34
3	23.54	23.96	25.05	18.14	20.95
4	25.23	25.21	26.35	18.80	22.70
5	27.15	26.56	27.75	19.48	24.57
6	28.83	28.03	29.25	20.18	26.58
7	29.69	29.63	30.86	20.90	28.71
8	30.31	31.39	32.61	21.65	30.96
9	33.76	33.34	34.53	22.43	33.32
10	35.24	35.50	36.64	23.24	35.79

.

The performance of the individual techniques is summarized in

Table 4. Appraisal Criterion of Proposed ANFIS LOL Model.

Metric	MAD	RMSE	MAPE
ANFIS Model	0.487	0.332	1.822
Chendong [31]	1.252	1.851	4.726
Vaurchex [32]	7.778	66.809	27.336
De Pablo [33]	1.859	4.495	7.457

.

It can be conclusively underlined that the prediction capability of the intelligence technique is superlative to all the existing methods in the context of the MAD, RMSE and MAPE.

5. Discussion

The surveillance of transformer conditions is conducted routinely to institute the health condition of transformers. In conformity with the results of such surveillance, remedial measures can be taken to safeguard and protract the lifespan of the unit in service. The transformer dielectric materials are composed of organic compounds, and they are therefore susceptible to decomposition at high thermal loading conditions.

When dielectric oil has combined chemically with oxygen to a particular extent, it will heretofore have fulfilled its intended purpose, and in practice, oil regeneration and change of oil will be an option for the transformer owner [8,9]. Correspondingly, cellulose paper is also susceptible to thermal ageing during service. Nevertheless, considering that the same maintenance scheme does not apply to cellulose and the ageing mechanics of cellulose decomposition are irrevocable, on that account, it concludes from this that the lifespan of a unit in the fields is determined by the lifespan of the cellulose paper. It is essentially unpragmatic to evaluate the condition of the cellulose paper rigorously aside from draining oil, removing the steel tank top cover of the transformer and extracting a physical sample. Lamentably, this is impracticable in the field; consequently, the status of the cellulose paper has to be evaluated implicitly. As discussed in Section 2, furan compounds are the by-products of the ageing mechanics of cellulose, and the quantity disintegrated in the dielectric oil can be harnessed as an index of the extent of decomposition of the cellulose. Notwithstanding, in order to associate the furan content, specifically 2FAL, of a unit in the field to the magnitude of cellulose degeneration, consumed life and consequently the remnant life cycle, more investigation and laboratory examinations are required.

The proposed artificial intelligence technique incorporating fuzzy logic and ANFIS in this work has not been explored previously, which makes it imperative in advancing the existing DP models [31,32,33] discussed in Section 3.

In

Table 5. Comparison of present study to previous investigations.

Model	Contribution
Chendong Model [31]	In these models, the DP is estimated using the measured 2FAL collected from a fleet of transformers at specific regions and the LOL is estimated using a separate Equation, i.e., Equation (4).
Vaurchex Model [32]
De Pablo Model [33]
Proposed ANFIS Model	In the present study, the remnant cellulose life of the transformer is achieved from an uncommon ANFIS modelling algorithm, and its performance has been studied against models above. There is currently no existing literature that covers the prediction of remnant life using 2FAL and DP. ANFIS offers significantly better learning capability: for the same network intricacy, a significantly lower MAD, RMSE and MAPE compared to existing models, as shown in Table 4. Essentially, the proposed approach independently predicts the LOL using the measured 2FAL and DP samples.

, the contribution and improvements of the current research in comparison with previous research works has been tabulated.

6. Conclusions

In the present study, an artificial intelligence technique incorporating fuzzy logic and ANFIS has been proposed to predict the remnant life cycle of the transformer by using the degree of polymerization of the cellulose insulation and the concentration of 2-Furfuraldehyde measured from the transformer oil samples. In addition, the study has benchmarked this proposed method by employing existing approaches to predict the transformer remnant life cycle. These approaches have been acknowledged to be pertinent in predicting the transformer life cycle; however, the prediction capability of the intelligence technique is superlative to all the existing methods in the context of the correlation coefficient and in comparison, with measured data obtained. The ANFIS model has been ascertained to provide an accuracy of 99.86%, for all the testing data. This underscores the model’s prediction dominance and can hence forth contribute to eluding the deficiencies of the PD and LOL models in the literature in predicting the remnant life cycle of cellulose insulating material in power transformers.

By examining the impact of 2FAL and DP present in the oil and cellulose, respectively, this work established that physically measured DP undoubtedly gives a significant rise to the decomposition of cellulose and ultimately the remnant loss of life of a unit in service, with a strong correlation between the amount of 2FAL concentration in the oil and DP on the cellulose and the remnant lifespan of the unit. Nevertheless, the potency of this impact has been demonstrated by the application of ANFIS after carrying out laboratory and site measurements.