How to Manage Electricity Costs in Laos with AI — Leveraging EDL Meter Reading Data

Introduction
By organizing EDL meter reading data in spreadsheet software and combining anomaly detection with consumption forecasting, you can identify, in concrete numerical terms, specific improvement points that lead to reduced utility costs in Laos. This article is written for facility managers and business owners who operate factories or hotels in Laos and are looking for ways to reduce their EDL electricity bill burden. By the end of this article, you will be able to carry out, on your own within your organization, the full four-step process: preparing and organizing meter reading data, analyzing usage patterns through statistical threshold judgment and forecast calculations, implementing improvement measures, and verifying results. We also cover specific topics such as data acquisition channels, tools needed for analysis, guidelines for investment decisions, and how to avoid common pitfalls.
AI analysis of electricity costs can only get started once you have the raw data in hand—namely, EDL meter reading data. Having your customer ID, meter number, and past billing data ready, and establishing in advance your acquisition channel and analysis setup, will determine the accuracy of all subsequent steps.
How to Obtain EDL Meter Reading Data
At present, the reliable channels available are retrieving history through the official EDL mobile app or from PDF invoices. The EDL app is available on both Google Play and the App Store, and once you register your customer ID and meter number, you can check your consumption history and pay bills. Check the monthly usage (kWh) and billed amount one by one from the history screen within the app, and if possible, export the past 12 months' worth of data into a spreadsheet so you can later compare seasonal variations during analysis.
If the history display in the app is insufficient, a practical approach is to manually record the meter number, reading date, usage, and amount listed on paper or PDF invoices. For facilities with a large volume of invoices, you could also consider scanning invoices on a monthly basis, extracting the figures with OCR, and then visually verifying them.
Another channel worth mentioning is the introduction of smart meters along with a HES (Head-End System) and MDMS (Meter Data Management System). This is a mechanism sometimes mentioned in news coverage of EDL's distribution network development or in international organizations' support projects, and if adoption progresses, it may become possible to obtain data at finer intervals for target regions or large-volume contract customers. However, the author has not been able to confirm any official announcement from EDL or related government ministries regarding a specific rollout schedule or scope, so this remains speculative at this stage. Whether your own meter is covered by such a system needs to be confirmed directly with EDL. For the time being, it is realistic to focus primarily on data acquisition through the app and invoices, and add smart meter–based data once it becomes available.
IT Tools and Skills Required for Analysis
The tools you need vary greatly depending on facility size and data acquisition channel. At the stage of gathering monthly data from the app or invoices, spreadsheet software with pivot aggregation is sufficient for analysis. With a spreadsheet that has columns for meter number, reading date, usage, and amount, you can go as far as tracking monthly usage trends and identifying peak months without any specialized tools.
If, in the future, you become able to obtain interval data (usage in 15-minute or 30-minute increments) via smart meters, the volume of data you handle will increase dramatically. At that stage, you will need a time-series database to accumulate the data, a BI tool for dashboard visualization, and an analysis environment (such as a statistical processing environment like Python or R) to run anomaly detection and forecasting models. On the distribution side, the HES and MDMS are responsible for quality-checking the raw data and standardizing its format, so the data your facility receives is expected to already be in an organized form.
The criterion for judging which stage you are at is the time granularity of the data on hand. As long as you only have one data point per month, a spreadsheet will suffice, but once daily or hourly figures start increasing, creating pivot tables will begin to take considerable time. That point is a good indicator for when to introduce an analysis environment.
In terms of skills, monthly aggregation and chart creation can be handled by facility managers themselves, but running algorithms for consumption forecasting or anomaly detection requires knowledge from a data engineer or analyst. If you don't have a dedicated specialist in-house, a reasonable approach is to start with spreadsheet-based visualization and consider external analytical support once the data volume increases.
Step 1: How to Organize Meter Reading Data?

Once monthly data has accumulated, before analysis you must align the granularity and fields of the raw data on a per-meter basis. If inconsistencies in how dates, usage, and amounts are recorded are left unaddressed, the figures will fail to match up at the aggregation and visualization stage.
Standardizing and Cleaning Data Formats
To eliminate inconsistent notation, the starting point is to decide on a fixed column structure in the spreadsheet before entering data. The minimum required columns are four: meter number, reading date (year-month-day), usage (kWh), and billed amount (LAK). Since date formats may differ between the app's history screen and paper invoices, align them to a single format such as "YYYY-MM-DD" at the time of entry, and make sure full-width and half-width numerals are not mixed.
Missing data and duplicates also need to be addressed. If there is a month for which no invoice can be found, do not leave the cell blank—clearly mark it as "missing" so that it is excluded from the average when later aggregating. If two rows end up being entered for the same meter number, cross-check the billing date and amount to determine which one is correct before deleting the other. For rows containing abnormally large or small values, check against the original invoice to determine whether it is a data entry error or an actual change in usage; if it is an error, correct it, and if it is an actual change, leave a note and include it in the aggregation.
Units of measurement should also be unified. Standardize usage in kWh and amounts in LAK, and rather than embedding comma separators or unit symbols (kWh, LAK) within the numeric cells themselves, place them in separate column headers—this makes automatic aggregation easier in later steps. If you are managing multiple facilities or multiple meters, adding a facility ID column so that rows can be sorted by facility name or meter number will make it easier to narrow down the aggregation scope. While this work takes some effort, if the data volume is around one entry per month, it can be handled entirely within the scope of spreadsheet software.
Visualizing Usage by Time of Day and Month
Once the column structure has been organized, converting the usage figures into a graph makes it easier to spot where waste is occurring. At the stage where only monthly data is available, start by creating a line graph with months on the horizontal axis and usage (kWh) on the vertical axis, lining up 12 months' worth of data on a single chart. This alone can reveal differences between the dry and rainy seasons, and overlaps with months of heavy operation for manufacturing businesses or months with many guests for hotel businesses.
Using a pivot table, you can automatically aggregate monthly usage by facility ID or meter number. If you are managing multiple buildings or meters, placing facility ID in the pivot's rows, month in the columns, and usage (kWh) in the values lets you compare at a glance which facility is increasing its usage and during which period. If you also arrange billed amounts (LAK) in the same format, you can compare whether changes in usage match changes in amount.
Pay attention to graph scale as well. When overlaying multiple facilities on a single graph, the line for a large-scale facility may stand out so much that fluctuations at smaller facilities get buried. When there is a large difference in scale between facilities, prepare a separate graph that converts usage into year-over-year percentage change, so that the magnitude of fluctuations can be compared regardless of scale.
Once daily and time-of-day data becomes available via smart meters in the future, add a line graph dividing a day into 24 hours, and a bar graph by day of the week. This reveals whether usage remains even outside operating hours, and whether peaks are concentrated during the day or at night, providing a basis for judgment in the next analysis stage of detecting anomalies and peaks.
Step 2: How to Analyze Usage Patterns with AI?

After grasping trends through monthly line graphs and pivot aggregations, the next step is to separate "unusual movements" and "future usage" out of the numbers. Unlike descriptive aggregation, this stage uses threshold judgments and comparative calculations against past data to numerically pick out waste that is hard to notice by eye alone.
Methods for Detecting Anomalies and Peak Usage
By setting a threshold and building a mechanism to automatically extract months or days that exceed it, anomalies that would be overlooked visually can be surfaced numerically. At the monthly data stage, a manageable starting point is to calculate the average and standard deviation from the past 12 months of usage (kWh), and flag months that significantly exceed the average as anomaly candidates.
Calculation example: Find the average and standard deviation of usage over the past 12 months, and treat any month exceeding average + 2 standard deviations as an anomaly candidate. If such a month actually occurs, go back to the original invoice and cross-check whether it is a reading error, an increase in the number of units in operation, or a change in air conditioning settings, in order to isolate the cause.
In a spreadsheet, this can be operated simply by adding one column that uses an IF function to flag threshold overruns.
# Pseudocode (check your spreadsheet software's help for exact function names and arguments) = IF(usage > average + 2*standard deviation, "anomaly candidate", "normal")
At the stage where daily and time-of-day data is available, check whether a constant value continues during hours outside operation (late night, days off) when usage should be close to zero. This indicates a state where standby power from air conditioning or equipment is lingering; setting up extraction conditions specifically for "time periods that should normally be close to zero," separate from threshold judgment, allows you to find waste that peak detection alone cannot catch.
Note also that the threshold value is not fixed and should be reviewed seasonally. For facilities where air conditioning load changes between the dry and rainy seasons, using a single average and standard deviation for the whole year increases false detections, so it is practical to calculate data separately for the dry and rainy season periods and maintain period-specific criteria for judging anomaly candidates. Providing a field to record the judgment date, cause, and outcome for rows that exceed the threshold will shorten the review work in subsequent months.
Consumption Forecasting Models by Equipment and Time
A predictive model that estimates next month's or next week's usage from past usage data should be built as a separate calculation from anomaly detection. At the stage where only monthly data is available, a manageable method is to multiply the moving average of the most recent 3 to 6 months by the ratio to the same month of the previous year; this can be calculated in a spreadsheet without needing any special analytical environment.
Calculation example: If the average usage over the most recent 3 months is 3,400 kWh, the same month last year was 3,000 kWh, and the same month two years ago was 2,800 kWh, apply the year-over-year growth rate (approximately 7%) from the same month last year to the most recent 3-month average, giving a forecast of approximately 3,640 kWh for the next month. Compare this forecast with the actual billed amount every month, and if there is a large discrepancy in any given month, check for changes in operating conditions (introduction of new equipment, fluctuations in room occupancy rate, etc.).
If you want to break down forecasts by equipment, an effective method is to calculate estimated usage for major equipment such as air conditioners and boilers based on operating hours and rated output, then take the difference from the actual measured value for the meter as a whole. The larger this difference, the more likely it indicates hidden, unaccounted-for consumption sources such as equipment outside the measurement scope or wiring losses.
At the stage where daily and time-of-day data is available, regression analysis incorporating day of the week and time of day as variables can improve forecast accuracy for each operating pattern. For manufacturing businesses, adding production line operating shifts as an explanatory variable, or for hotel businesses, adding guest count and check-in time slots, makes it easier to capture fluctuations beyond seasonal factors compared to a simple moving average. If there is equipment where the gap between forecast and actual values is not temporary but persistent, suspect equipment degradation or a setting error, and record it as a candidate for the next round of improvement measures. Aim to review the forecast about once a month, and update the model's assumptions whenever operating conditions change significantly.
Step 3: How to Develop Improvement Measures from Analysis Results?

Only once anomaly candidates and forecast deviations have accumulated do you finally have the material needed to judge which equipment to fix and which investments to commit to. Simply looking at the numbers does not lead to cost reduction; you need to separate the target of action and the scale of investment according to the type of cause.
Key Points for Reviewing Equipment Operations
The causes behind anomaly candidates and forecast deviations require different responses depending on the type of equipment involved. By first sorting the causes into three categories, it becomes clear where to start.
| Type of Cause | Specific Symptom | Target for Review |
|---|---|---|
| Residual consumption outside operating hours | A constant level continues even at night or on non-business days | Auto-shutoff settings for air conditioning/lighting, cutting off standby power |
| Concentration during peak hours | Usage spikes only during specific time periods | Spreading out operating schedules, peak shifting |
| Decline in efficiency of individual equipment | Usage increases even though operating hours remain unchanged | Checking power factor, inspecting/replacing aging equipment |
Residual consumption outside operating hours can sometimes be resolved simply by reviewing the timer settings for air conditioning or lighting. Since this can be addressed without additional investment, it is a suitable first item to check.
Concentration during peak hours is addressed through changes to the operating schedule, such as shifting start times or spreading out equipment operating hours. Since this does not require changing the equipment itself, it is worth trying as a preliminary step before making investment decisions.
A decline in efficiency of individual equipment is a case where a drop in power factor or aging is suspected. In this case, inspection, parts replacement, or in some cases equipment replacement becomes necessary, making this the target that moves on to the next step of estimating return on investment. Keeping records that separate out the type of cause means the material for judging how much to spend on which measure carries directly over to the next process step.
How to Estimate ROI
Investments addressing a decline in efficiency of individual equipment should be judged only after the payback period is expressed numerically. The most manageable indicator is the simple payback period, which can be calculated simply by dividing the investment amount by the annual savings.
Calculation example (these are hypothetical values for explanatory purposes; please check actual rates against EDL's published rate schedule): Suppose replacing an aging air conditioning unit costs 30 million LAK, the monthly usage reduction after replacement is 1,200 kWh, and the unit price is assumed to be 5,000 LAK/kWh. The annual savings would then be 1,200 kWh × 5,000 LAK × 12 months = 6 million LAK. In this case, the simple payback period is 30 million ÷ 6 million = 5 years. If this number of years is shorter than the equipment's remaining useful life, it becomes easier to decide to proceed with the investment; if it is longer, the decision tends to lean toward prioritizing life extension through partial inspection rather than replacement. When actually calculating this, check the EDL published unit price corresponding to the target facility's industry classification and consumption tier, and substitute that value in.
When examining multiple improvement measures at the same time, comparing not just the simple payback period but also net present value (NPV) or internal rate of return (IRR), which account for the discount rate, makes it easier to prioritize among measures of differing investment scale. Since calculating NPV or IRR requires assumptions about future rate levels, check the current rate structure against information published by EDL or the relevant ministry before applying the figures. During periods when a rate revision is scheduled, calculating under both the pre- and post-revision assumptions and confirming that the payback period stays within an acceptable range under either case helps avoid having to backtrack on an investment decision.
For cases where judgment is difficult, it is a sound approach to lower the priority by one level, first implement no-investment measures such as addressing residual consumption outside operating hours or spreading out peak-hour concentration, and then recalculate the investment proposal using the actual measured values after the reduction. This avoids the mistake of deciding on an investment based on an assumption of an overly optimistic reduction effect.
Step 4: How to Implement Improvements and Verify Results?

Making an investment decision alone does not reduce costs. By checking numerically whether the implemented measure lowered usage by the intended amount, and building in a mechanism to adjust conditions if the effect is weak, you can ensure that improvement measures do not end as a one-time response.
Monitoring System After Implementation
From the month a measure is implemented, deciding on a single dedicated comparison standard for that measure keeps the effect assessment from becoming inconsistent. The standard to use is EnPI (Energy Performance Indicator), and what matters is that the indicator is "usage per unit of operation" rather than usage itself. For manufacturing, this means calculating kWh per unit of production; for the hotel industry, kWh per occupied room, and fixing the 3 to 6 months prior to implementing the measure as the baseline value.
Calculation example: If the baseline value before the measure was 120 kWh per ton of production, and the actual measurement for the 3 months after the measure is 105 kWh per ton, then the actual reduction rate, excluding the effect of fluctuations in operating volume, is about 12.5%. Comparing only absolute usage values mixes the effect of the measure with the impact of changes in production volume or room occupancy rate, so the practice should be switched to recording the value divided by EnPI each month.
The monitoring frequency should basically be once a month, and for investment projects addressing a decline in efficiency of individual equipment, record the assumed reduction amount used in the payback period calculation alongside the actual measured EnPI value. If the reduction rate is lower than expected for 2 to 3 consecutive months, suspect a reversion of setting changes or the emergence of a new anomaly candidate, and return to the anomaly detection process to recheck the cause. Preparing a record table with five columns—measure name, implementation date, baseline value, measured EnPI, and difference—makes it easier to trace the history when reviewing later.
For facilities with large seasonal factors, it is also effective to maintain separate EnPI baseline values for the dry season and the rainy season respectively. Evaluating the whole year with a single baseline value risks confusing changes in air conditioning load with the effect of the measure, mistakenly flagging a measure that is actually working for review. Separating the baseline values by period avoids this misjudgment.
How to Build a Continuous Improvement Cycle
If you finish with a single round of measures and monitoring, the same anomaly may recur at the same time the following year. By building in a mechanism that subjects the actual measured values accumulated in the EnPI record sheet to a higher-level review every six months, you can shift from a practice of stacking one-off responses to a sustained operation.
In the semiannual review, check the following three points. First, whether the reduction rate of the measures recorded in the record sheet is being maintained. Compare against past actual EnPI values to see whether the setting changes have reverted, and if a reversion is found, carry out the setting adjustment again and record the date of execution. Second, check whether any of the cases referred to the investment-return estimate as a decline in standalone equipment efficiency remain unaddressed. If the rate level assumed for the payback period has changed, update the figures using the current published information and reorder the priority for starting work. Third, review the anomaly-candidate thresholds separately for the dry and rainy seasons. At facilities where equipment has been added or operating patterns have changed, recalculate the average and standard deviation using the most recent period and replace the thresholds accordingly each time.
Setting up a forum where management or facility managers review this semiannual check prevents the judgment from stopping with individual staff members alone. By adding the check results to the record sheet and carrying them forward as the baseline for the next six months, you can reduce the effort of starting the review process from scratch each time.
What Are Common Pitfalls?

Cases where AI analysis of power costs fails to lead to cost reduction can be divided into two types of causes: failures where the quality of the raw data is low and the analysis results themselves cannot be trusted, and failures where the analysis results are correct but are never implemented on-site. Both can be prevented by not skipping steps and going back to confirm the preceding process.
Reduced Analysis Accuracy Due to Insufficient Data
If you try to build an anomaly detection or forecasting model before even six months of monthly data have been gathered, the calculation of the average and standard deviation itself becomes unstable, and the threshold judgment ends up off the mark. The "average + 2 standard deviations" used in judging anomaly candidates is pulled toward extreme values more easily when there are fewer data points, which can cause even months that are no different from usual to be flagged as anomaly candidates. The causes can be divided into two: input omissions and how missing values are handled.
An input omission is a state where part of an invoice cannot be found, leaving a gap in the monthly data that gets tallied anyway. If a missing month is left blank and included in the average calculation, that month's figure is treated as zero, and the average and standard deviation end up calculated lower than the actual situation.
In cases where missing values are mishandled, there are examples where a missing month is entered as "0 kWh," causing anomaly judgments from the following month onward to become overly sensitive. The fix is to exclude the missing month from the data range and recalculate the average and standard deviation.
Another easily overlooked cause is mixing the figures for a month in which a meter was replaced or a contract was changed directly into the time series as-is. In the month the meter number was switched, the reading period may be shorter than usual, and a simple comparison as if it were a full month can make the usage amount look unnaturally low. For this month, add a "reference value" note, exclude it from the tally, and fill in using the surrounding periods instead.
Until the data reaches 12 months' worth, treat the results of anomaly-candidate judgments as reference values only, confine verification of the underlying cause to checking against the original invoices, and avoid carrying misjudgments caused by insufficient data into the execution plan.
Improvement Measures Failing to Take Root On-Site
Even after deciding on setting changes or schedule distribution based on analysis results, there are cases where, once the staff member who implemented them is replaced due to transfer or resignation, things revert to the original settings. Changes such as air conditioner timer settings or staggered start times are the kind of change that an individual staff member can revert simply by personal judgment, so this is a measure that, if not documented, tends to be forgotten by everyone.
To prevent this recurrence, don't rely on verbal communication or individual memory for the setting-change details; instead, keep a written record of "the setting values themselves," separate from the EnPI record sheet. Compile the timer setting times, the start times changed for peak shifting, and the model numbers and work dates of equipment inspected or replaced onto a single sheet, and post it near the equipment or in the control room, so that the same state can be restored even when staff members change.
When the same equipment is operated across multiple shifts or multiple sites, a single sheet may not reach every staff member. Adding a setting-value confirmation column to the shift handover notes, with an item to cross-check the actual displayed value against the recorded value at shift change, can reduce the chance of settings reverting without anyone noticing.
Another reason measures fail to take root is that the figures showing the effect of the measure are not shared with on-site staff, so the meaning of the change never gets through. By feeding the reduction rate confirmed in the semiannual review back, in a simple form, to the department that handled the setting change, and showing how the change connected to the figures, on-site staff come to avoid reverting the setting on their own initiative. Establishing a practice where reverting a setting requires a supervisor's approval also prevents a change from being nullified by an individual's judgment alone.
Frequently Asked Questions

To supplement the steps covered so far, let's organize three points concerning data acquisition channels, analysis structure, and the approach by facility scale.
Q1: Where Can EDL Meter Reading Data Be Obtained?
At present, there are two reliable channels for obtaining data: the official EDL mobile app and paper/PDF invoices. With the app, once you register your customer ID and meter number, you can check consumption history and bill payments; with invoices, you manually record the reading date, usage, and amount. Data via smart meters, planned for introduction under the grid improvement project, is expected to become available in stages, starting with target areas and large-volume contractors, so not all facilities are covered. You will need to check with EDL to find out whether your own meter is included.
Q2: Is a Specialized Data Scientist Required for AI Analysis?
The answer depends on the stage of analysis. Up through monthly aggregation, visualization via pivot tables, and threshold-based flagging, facility managers themselves can handle this with spreadsheet software. Once you move on to consumption forecasting via regression analysis or advanced anomaly detection using interval data, you need the expertise of a data engineer or analyst. If there is no dedicated specialist in-house, it is not too late to consider external analytical support at that stage.
Q3: Can Small Facilities Also Start AI Management of Power Costs?
This can be started even at small facilities. Even with only one monthly invoice's worth of data, aggregation and threshold judgment using spreadsheet software are feasible, and the fewer the pieces of equipment, the faster the cause can be isolated.
A suitable approach for small facilities is as follows.
- Record usage and amounts for the past 12 months from the EDL app or invoices
- Grasp seasonal variation using monthly graphs
- Identify anomaly candidates using the average and standard deviation
- Start with measures that require no investment, such as eliminating consumption that continues outside operating hours
Even without finely spaced data like that from large-user smart meters, monthly data alone is enough to progress from identifying anomaly candidates to making no-investment improvements. At facilities with few pieces of equipment, once the cause of efficiency decline has been identified, the cost-benefit estimate for investment can also be calculated by narrowing the focus to just one or two pieces of equipment, so there is no need to force the scale of analysis to be larger than necessary.
Summary

The flow of analyzing EDL metering data with AI runs continuously from data preparation through effect verification. First, obtain the past 12 months of usage and amount data from the EDL app or invoices, align the columns for meter number, reading date, usage, and amount, and eliminate inconsistencies in notation. After visualizing seasonal variation with monthly and time-of-day graphs, identify anomaly candidates using the average and standard deviation, and forecast the following month's usage using moving averages or year-over-year comparisons for the same month. Classify causes into three types—consumption remaining outside operating hours, peak concentration, and declining equipment efficiency—start with measures that require no investment, and judge investment in efficiency decline by simple payback period. After implementation, compare usage per unit of operation against a baseline using EnPI, and review thresholds and the reduction rate of measures every six months, turning this into an ongoing operation rather than a stack of one-off responses. If you rush to make judgments with insufficient data, or return to the field without recording configuration changes, the results of your hard-won analysis will be lost at the implementation stage. Start by organizing the past 12 months of data into a single spreadsheet.
References
- www.nta.go.jp/about/organization/ntc/kenkyu/ronsou/111/04/04.pdfAccessed November 18, 2026
- samrepo.nlr.gov/help/revenue_ppa.htmlAccessed November 18, 2026
- www.meti.go.jp/policy/it_policy/jinzai/AIguidebook_DemandForecasts_FIX.pdfAccessed November 18, 2026
- www.enecho.meti.go.jp/category/saving_and_new/saving/enterprise/factory/report/…Accessed November 18, 2026
- cloud.google.com/blog/ja/products/management-tools/cloud-monitoring-adds-long-l…Accessed November 18, 2026
Author & Supervisor
Chi
Majored in Information Science at the National University of Laos, where he contributed to the development of statistical software, building a practical foundation in data analysis and programming. He began his career in web and application development in 2021, and from 2023 onward gained extensive hands-on experience across both frontend and backend domains. At our company, he is responsible for the design and development of AI-powered web services, and is involved in projects that integrate natural language processing (NLP), machine learning, and generative AI and large language models (LLMs) into business systems. He has a voracious appetite for keeping up with the latest technologies and places great value on moving swiftly from technical validation to production implementation.


