MACHINE LEARNING
SteamWise Predict
Enter boiler-plant measurements. Explore the classification predicted by a genuinely trained 100-tree Random Forest, with model probabilities and range checks.
Try fault prediction ↗
BUILTSMART AI CAPSTONE • ENERGY & INTELLIGENCE
The boiler is talking.
Are we listening?
Every building has a story hidden in its operating data. SteamWise AI explores how we can make that story easier to understand—and how responsible machine learning could help us listen.
Simulated hot-water boiler research · Human judgment always comes first
744kDevelopment observations
149kIndependent verification
6Operating-condition classes
12Interactive data investigations
01 / THE REASON
Not another black box.
A better way to ask questions.
Boilers, pumps and sensors generate signals continuously. Weather, occupancy and heating loads change what “normal” looks like. The opportunity is to combine operating evidence with sound engineering judgment—not replace it.
Where engineering meets dataTHE CORE QUESTION
Can a machine learn patterns of abnormal operation without mistaking a changing baseline for a fault?
That is the technical question behind SteamWise AI—and the reason external validation matters.
02 / TWO WORKING EXPERIENCES
Explore the technology.
Not just the slides.
COMPUTER VISION
SteamWise Vision
See how a component photograph can become an interactive Power Engineering lesson—with identification cues, safety learning notes and syllabus links.
Launch Vision ↗03 / THE SURPRISE
The model looked nearly perfect.
Then we changed the test.
Our final DP3 model posted 99.73% on an internal random split. When challenged against a separate simulated operating period, the score was materially lower. That gap became the most valuable result of the project.
DP3 RANDOM FOREST
Internal test accuracy99.73%
↓ Different test. Different conclusion.
External temporal accuracy89.28%
And for Normal operation, external recall was only 30.42%. A headline metric cannot tell the whole story.
Investigate what changed ↗
04 / A BIGGER IDEA
The model is the proof.
The learning loop is the product.
Imagine a future owner-specific system that learns from authorized building measurements, flags patterns for professional review, and adapts through supervised monitoring and controlled retraining.
Then connect that diagnostic capability to visual equipment education and, one day, an interactive Virtual Boiler Simulator.
05 / PURPOSEFUL IMPACT
More intelligence. More responsibility.
Affordable & Clean Energy
Explore the potential for more informed energy management and efficient operation.
Responsible Consumption
Support thoughtful resource use, equipment understanding and maintenance decisions.
These are prospective contributions, not demonstrated energy, cost or emissions savings.
DATA ANALYSIS / COMPONENT 2
Before teaching a machine,
understand the system.
Twelve questions. Six simulated operating conditions. Two data populations. One place to explore the evidence for yourself.
THE ORIGINAL INVESTIGATION
What the Power BI exploration revealed.
The source dataset models two boilers and two circulating pumps serving a simulated 12-storey Chicago office building. This original project screenshot preserves the exploratory context.
01
Structure matters
Temperatures, pressure, flow, gas and equipment states only make sense in the context of the plant.
02
Not all data are equal
Imbalanced conditions, repeated readings and near-constant signals require deliberate preparation.
03
Realism has limits
Simulated scenarios provide a controlled benchmark—not proof of real building performance.
INTERACTIVE ANALYTICS LAB
Ask better questions.
Discover deeper answers.
Every chart below uses source-backed aggregates or clearly identified sampled observations. Hover, filter, zoom and export chart images.
Reviewable evidenceDevelopment: 744,090 rowsVerification: 149,328 rows
Scope and class filters apply to source-distribution views; model-verification charts always use the independent external set. The monthly trend always shows the original fault-free scenario.
WHAT COMES NEXT
Visual exploration becomes model discipline.
EDA revealed why a robust data audit is essential before modelling—and why highly correlated simulation features can create misleading confidence.
Follow the ML pipeline ↗THE DETECTIVE STORY
We built a model.
Then we challenged it.
Seven steps in an applied research journey: more questions, better data understanding, and a hard-earned respect for independent evaluation.
THE PIPELINE
A disciplined journey from observation to evidence.
Raw baseline
20 candidate predictors after excluding problematic pump-power signals.
DP1 audit
Missingness, repeated states, class balance, unusual values and potential shortcuts.
DP2 clean
Remove two constants and an almost-constant setpoint; retain 17 predictors.
DP3 engineer
Add three ΔT features, combined boiler flow and combined gas input.
Train
RandomForestClassifier, 100 trees, six operating categories.
Verify
Evaluate against 149,328 independent Nov–Dec simulation records.
Interrogate
Investigate normal-class recall, confusion pairs, bias and deployment limits.
THE EXPERIMENT
Three iterations. One decisive test.
Raw achieved the strongest external numerical score, but retained possible simulation shortcut behaviour. Physically meaningful DP3 features improved interpretability without improving temporal verification accuracy.
RAW / EXTERNAL93.67%
Accuracy
DP2 / EXTERNAL89.96%
Accuracy
DP3 / EXTERNAL89.28%
Accuracy
DP3 / INTERNAL99.73%
Random 80/20 accuracy
THE REAL LESSON
Generalization beats a headline score.
Against genuinely later simulated data, the model's results were weaker than its internal random-split performance suggested.
Accuracy alone also hid an important weakness: DP3 recognized only 30.42% of actual Normal observations.
Explore all external metrics ↗PRIMARY EVIDENCE
See the workflow, not just the conclusion.
ENGINEERING INTERPRETATION
Better feature logic does not guarantee a better score.
Feature engineering makes a model more interpretable, but operational change, label imbalance and possible simulation artifacts can still undermine generalization.
Next research priorities include time-aware cross-validation, investigation of shortcut predictors, improved Normal recognition and fault-onset labels suitable for actual early-warning evaluation.
Try the real model ↗
WORKING APPLICATIONS
Experience the tools.
Challenge the technology.
One system interprets operating measurements. Another helps students explore equipment through images. Both require informed human judgment.
01 / STEAMWISE PREDICT
What do these measurements suggest?
Submit 17 simulated-process measurements. The server calculates five engineering features and returns an actual prediction from a trained 100-tree Random Forest.
Checking prediction engine…
◉
Your prediction will appear here.
Choose a sample state, modify any variables and select Analyze operating pattern.
MODEL OUTPUT · SIMULATED DATA
Top-class model score
This classifier has not been validated for real facilities. Probabilities are uncalibrated model estimates—not safety assurance or proof of a fault.
THE ACTUAL INFERENCE MODELRandom Forest · 100 trees
Portable DP3 deployment model, separately trained from academic mlOS. Independent verification: 89.48% accuracy and 82.96% macro F1 on 149,328 simulated observations.
02 / STEAMWISE VISION
See a component.
Discover its story.
A separate Google AI Studio multimodal prototype helps students connect equipment photographs with Power Engineering knowledge, visual identification cues and safety-learning notes.
MULTIMODAL AI LEARNING
Turn equipment into an interactive lesson.
Explore low-water fuel cut-off devices, pressure gauges and component functions. Visual recognition is educational; it does not confirm equipment condition.
Launch SteamWise Vision ↗Google AI Studio may require sign-in. Do not submit sensitive facility or personnel images without authorization.
Low-water fuel cut-off
Explore equipment identification, a protective function and its difference from routine feedwater control.
Explore in SteamWise Vision ↗Pressure gauge
Explore local indication and how it differs from automatic pressure control and independent protection.
Explore in SteamWise Vision ↗03 / THE FUTURE
From two capabilities to a Virtual Boiler Academy.
A future simulator could combine engineering-based scenarios, machine-learning explanations and visual component education. This is a concept—not a completed or certified operator-training system.
01Simulate
Explore fictional plant operating changes in a controlled environment.
02Interpret
Compare physical cause-and-effect with data-driven classification.
03Learn
Understand equipment, protective functions and evidence-based decisions.
THE PROJECT LIBRARY
Everything behind
the story.
Watch, read, inspect, download and interact. A curated evidence shelf for industry, education and technical reviewers.
NOTEBOOKLM VIDEO / PUBLISH-READY SLOT
Watch the story. Then explore the evidence.
The NotebookLM executive video is currently in preparation. This section is fully designed for either direct MP4 upload or a public, embeddable video URL.
▷
STEAMWISE AI · VIDEO OVERVIEW
A story worth watching.
Video coming soon · Upload or paste its URL in site-config.jsonVideo embedding never relies on a fake preview. If an external host blocks embedding, the website will show a verified "Watch video" button instead.
DOWNLOAD & EXPLORE
Choose your level of detail.
01 / READ
Capstone report
The complete project story, methodology, results and responsible AI framework.
Download Word ↗02 / VIEWExecutive presentation
12-slide story for industry and policy stakeholders.
View PDF ↗03 / PRESENTEditable slides
Original presentation assets ready for PowerPoint.
Download PPTX ↗04 / EXPLOREInteractive analytics
Five executive and seven advanced questions, backed by the project data.
Open dashboard ↗05 / DOWNLOAD20 statistical exhibits
Original 300-dpi research graphics with source-driven comparisons.
Download ZIP ↗06 / AUDITMaster reference
Extended project compendium and technical context.
Download Word ↗07 / SOURCELBNL dataset
Original simulated boiler plant dataset and documentation, DOI 10.25984/1881324.
Open official source ↗08 / TRYSteamWise Vision
Multimodal component-identification and Power Engineering study prototype.
Launch Vision ↗09 / RUNPrediction engine
Actual server-based trained model inference, with warnings and plain explanations.
Explore predictions ↗ASK THE PROJECT
Have a question? Ask SteamWise.
The Jotform AI assistant is connected to the project knowledge base. It can help navigate concepts, figures and resources; verify important answers against primary evidence.
SteamWise AI Assistant
Try: "Why did DP3 accuracy decrease during temporal validation?"
Or open full-screen ↗The chatbot is provided by a third party. Clicking “Embed” opens its content within this page and may enable third-party processing.
RESPONSIBLE USE & PRIVACY
AI informs.
People decide.
Prediction quality, privacy, technical boundaries and independent protection belong in the same conversation.
OUR FIRST RULE
Strong decisions require more than an algorithm.
SteamWise is a controlled simulation-based research exercise. It has never been qualified for industrial or commercial plant control, failure prevention or safety-critical diagnostics.
Authorized data
Public simulated records only. Future owner-data integrations require permission, controlled access and data-handling agreements.
Traceable transformations
Keep original data, audit DP1–DP3 operations and verify derived quantities before use.
Class-level accountability
Show Normal recall and misclassification patterns, not only headline accuracy.
Human oversight
Professional operators and engineers evaluate all physical implications and interventions.
Security & privacy
No raw personal data is required by the current prediction engine. Keep backend models and secrets off the public static assets.
Continuous review
Future monitoring must detect operating-data drift, verify changes and enforce human approval of retraining.
DISCLOSURE
What this experiment can—and cannot—prove.
What we demonstrated
- Reproducible simulated-data preparation and feature design
- Trained six-class Random Forest classification
- Independent temporal evaluation and source-backed graphics
- A working prediction interface and visual-learning prototype
What remains unproven
- Recognition reliability in other real buildings or climates
- Validated early fault-onset detection or lead time
- Real-world energy, cost or emission savings
- Safety integrity, operational diagnosis or autonomous control
CONTROLS ALONG THE PIPELINE
Build the safeguards into the workflow.
DATA
Permission, completeness, provenance and fairness across represented operating states.
MODEL
Feature integrity, reproducible tests, leakage assessment and per-class evaluation.
DEPLOY
Secure access, restricted functionality, transparent uncertainty and professional review.
EVOLVE
Ongoing drift checks, incident learning, versioning, audit records and human approval.
The original U.S. DOE dataset is publicly accessible. Check applicable licensing and attribution terms before redistributing original source material. This website links to the authoritative dataset rather than distributing the multi-gigabyte raw collection.
THE HUMAN BEHIND THE PROJECT
Engineering experience.
Human curiosity.
A project at the intersection of complex energy systems, responsible artificial intelligence and the people who operate them.
ABOUT THE DEVELOPER
Juan M. Galarza
Mechanical Engineer · PMP® · Energy & Asset Management · Applied AI
Over more than 25 years, I have worked across complex energy projects and operating assets in multiple countries, connecting engineering, operational discipline, project delivery and people.
SteamWise grew from a question that sits at the heart of my experience: How can technology help professionals understand equipment better and make more informed decisions?
The project combines a disciplined machine-learning exercise with a vision for practical, human-centred learning tools. It is one step in a much larger journey of energy, operational excellence and digital transformation.
WITH GRATITUDE
Built with guidance. Grounded in experience.
SteamWise AI was developed by Juan M. Galarza under the mentorship and supervision of Fariha Iffath as a capstone for the BuiltSmart AI program. Program delivery through Manpower Alberta, applied AI training by Braintoy, and funding by the Government of Alberta.
BUILTSMART AIMANPOWER ALBERTABRAINTOYGOVERNMENT OF ALBERTA
Institutional references acknowledge the educational program; they do not imply endorsement or certification of this software. Official logos can be added after brand permission and source verification.
THE LONGER HORIZON
From an experiment to a new learning environment.
The ambition is to explore site-specific diagnostics, computer vision for technical education and a future Virtual Boiler Simulator. The path forward will be guided by stronger validation, real-world data permissions and the expertise of the people who understand the equipment best.
Explore the technology ↗