9905-867 The good news is that with the latest technology, we are able to help operators optimise processes and minimise carbon emissions. In one recent use case, Schneider Electric deployed a customized, near-real-time machine learning model to monitor six sources of carbon emissions in a vacuum distillation unit, achieving carbon reduction goals.
Vacuum distillation units are widely used in different industries such as chemical and pharmaceutical production, crude oil refining, essential oil and fragrance manufacturing, food processing, production of hot base water required for ultra-pure water or desalinated water. The machine learning model built by Schneider Electric uses the AVEVA PI connector to analyze the data stream every five minutes to generate timely feedback on potential deviations in CO2 emissions. This enables operators to react quickly, investigate root causes, and make targeted adjustments to optimize processes and minimize CO2 emissions.
The above model is not only applicable to vacuum distillation units, but can also be migrated to d9905-867 ifferent industrial processes, thereby reducing environmental impact while improving operational efficiency and helping industry move towards a more sustainable future.
Using machine learning to predict carbon emissions
To achieve near-real-time CO2 tracking, the basic steps include validating operational data, determining emission benchmarks, using machine learning (ML) algorithms to predict emissions, flagging events in different operating states, and performing root cause analysis. During the project execution phase, project team experts will assist with the validation and correction of operational data, as well as provide process interpretation. Data scientists then focus on Feature Engineering, selecting machine learning algorithms, and determining metrics.
Ultimately, machine learning algorithms can predict key operating parameters based on specific fa9905-867 ctory operating conditions.
In Figure 1 (above), outliers based on vacuum feed and gas in the burner are initially identified. An outlier is an observation that has an abnormal distance from other values in the data set, shown as a purple line with a value of 1. The normal value index is the typical observed value in the data set, expressed by the value 0.
Then, after removing the outliers from the historical data, the ML model is trained based on the cleaned data, and the key operational parameters are predicted every five minutes through the ML model. In Figure 2 (below), some of the predicted KPI key performance indicators closely match the measurements, indicating normal operation, while others show significant deviations. These actions help us anticipate potential problems.
Data drift is also monitored in Figure 2, reflecting changes in statistical attributes over time, and evaluated using the Area under the Curve (AUC) indicator. Where, AUC close to 0.5 indicates minimal drift, close to 1 indicates more significant drift, and JS Divergence (Jensen Shannon Divergence) is used to measure the effect of drift on model performance. These evaluations help ensure that the model remains accurate and reliable as operating conditions change over time.
Use machine learning to find biases
In Figure 3, the ML model identifies the key factors that affect the target outcome in order to perform a root cause analysis of the bias. Provide insight into emission control decisions by constantly updating and ranking important features in real time. This number indicates the importance of a feature, and the greater the value, the greater the impact.
The figure also shows the mean, minimum, maximum, and trend of feature importance over time. With this data, we can intervene in a timely manner and seize opportunities to improve process control, performance and emissions reduction.
Integrating advanced machine learning models with the AVEVA PI System operational Big Data management platform enables organizations to maximize the potential of their operational data. As shown in Figure 4, this integration provides actionable insights to optimize device performance and enable data-driven decision making. By using models analyzed from historical data, businesses can make real-time predictions, detect biases and potential root causes, and thereby improve performance, reduce costs, and gain a competitive advantage.
The integration process is simple and easy to operate, and can be completed in the following few steps:
1. Configure a VM or cloud environment.
2. Configure PI system to realize real-time data storage and notification management;
3. Configure the Python environment and create the necessary files.
4. Set up PI connectors for common file and stream loaders to import external source data directly into the AVEVA PI System operational Big Data management platform.
