Demonstrating Machine Learning in Meter Reading

Overview

The figure shows the overview of ML processing for implausible reading results in SAP S/4HANA Utilities.

ML makes use of a large volume of historical utilities principal and transaction data for making forecasts, all of which are kept in the customer's system. ML uses both training and apply datasets in order to reach ML-relevant data for excluded billing documents. This information sets the foundation for figuring out the release confidence value for specific billing documents. Furthermore, this data is required for training a forecast model in ISLM.

The ISLM is enabled to define, train and activate a predictive model for the predictive scenario UTI_BI_OUTSORTED.

The report REML_BILLING_OUTSRTD_RELPRED selects all outsorted billing documents according to the selection parameters outlined by the user. It then calculates the release confidence value for each outsorted billing document in the selection, given there is enough historical data. Automatic release of outsorted billing documents can occur, subject to a defined threshold and the calculated release confidence.

Flowchart illustrating the process flow for training, applying, and displaying machine learning models in SAP utilities. The Train phase involves the ISLM Scenario and Model with UTR_MR_IMPLAUSIBLE, connected to CDS View Training and Customizing Timeframe. In the Apply phase, reports determine release confidence and automate releases, using Utilities Master Data and Utilities Transaction Data. The Machine Learning Processing Data links to these processes. In the Display phase, the SAP Fiori App resolves implausible reading results, supported by a CDS Model providing machine learning data.

The report results are stored in a database table and can be displayed in the application log:

billing document
release recommendation
indicator about whether the documents have been automatically released

The report needs the following prerequisites:

The training and prediction timeframe must have been maintained in customizing.
A trained and activated model version for the ISLM predictive scenario UTI_BI_OUTSORTED must be available.
If billing documents are to be released automatically, a release threshold of at least 0.5 must be provided on the selection screen.

The SAP Fiori app Resolve Outsorted Billing Documents is used to display the release confidence for each outsorted billing document, if the release confidence was correctly determined and the outsorted documents have not been released automatically.

The closer the release confidence value is to 100%, the greater the assurance of the related ML prediction that the document can be released.

ISLM Scenario and Model

The figure shows the ML model training based on historical data and the label determination for implausible meter reading results in ISLM.

The predictive scenario, named UTI_MR_IMPLAUSIBLE, functions as an SAP template that enables ML to manage uncertain meter reading results. This scenario encompasses an algorithm, a training dataset, an apply dataset, and a target variable, all of which assist in training the customer-specific predictive model. Both the training dataset and apply dataset contain the relevant features and labeling logic at the model level designed for unlikely reading results.

Initially, the predictive model is untrained, or in simpler terms, there is no existing model version. For usage, it is essential that the predictive model undergoes training and activation within the customer's system.

Flowchart illustrating the processing of Implausible Reading Results Historical Data. Starts with Reading Result Unprocessed, identified as Result Not Predictable and not relevant for model training. Other outcomes become Reading Result Corrected or Released, deemed Result Predictable and relevant for training. Criteria include Status Change (Implausible to Plausible or Released), Consumption Characteristics, and Attributes like Meter Reading, Meter, Register, and Installation. A decision point determines if results meet criteria, leading to a Label of either No Release or Release Document.

The predictive scenario is linked to the application using the ISLM and report REML_MR_IMPLSBL_RELPRED, so that the predictive results can be consumed by the end user in their own environment.

In addition to the agent's decision to release, correct or reverse a reading result, possible subsequent actions are taken into consideration to determine the expected label.

The result is a release confidence value, which can be used either to process the implausible reading result manually (label No Release), or to automatically correct, release or reverse it without further user interaction (label Release Document).

Training Period

The customizing table TEML_IMPLMRPARAM provides the training and prediction timeframes for implausible reading results.

Timeline diagram showing the training and prediction periods. The Training Start Date is 03/01/2017, and the Training End Date is 01/01/2019, covering a 24-month span. The Prediction Period extends beyond the Training End Date and continues until the current date of 03/01/2019. A 2-month gap exists between the end of training and today, highlighting the timeframe used for model predictions.

In this example,

the oldest document (actual reading date) considered for training dataset is dated 03/01/2017 which is 24 months earlier than today and
the last document considered for training dataset is dated 01/01/2019 which is 2 months earlier than today.

The timespan between 01/01/2019 and today 03/01/2019 indicates the offset timeframe, during which possible actions might be performed for released documents, meaning that the reading results within this timeframe should not be used for training.

According to the logic for the prediction period (apply dataset), no forecast is created for meter reading documents with an actual meter reading date that is earlier than the start date of the training. All other documents will be predicted.