Energy cost optimisation with AI, without the cloud
Summary
Energy cost optimisation with local AI: energy management software finds the saving potential in OPC UA data without the cloud. Try DataPortia free for 30 days.
Picture the situation: the energy cost per unit of product at an industrial plant has risen three per cent in six months. Nobody knows why. Total consumption looks normal on the monthly invoice, production volumes are unchanged — yet the cost keeps climbing. What leads to this is monitoring at too coarse a level.
The problem is not energy itself. The problem is that energy data is not interpreted in its process context. Once load, recipe, time and weather are brought into the analysis, the cost increase typically narrows down to 2–3 process stages or pieces of equipment — and those can be addressed. This calls for energy management software that combines OPC UA data with local AI.
In this article I go through why energy cost optimisation fails without process-level data, how local AI finds the saving potential — and why an on-premises model is the only realistic option for many industrial plants.
Why does energy cost optimisation fail without process-level data?
For many industrial plants energy is the single largest variable cost — often 20–40% of total production costs. Yet energy decisions are still made at too coarse a level: from the total figure on the monthly invoice, or from a single main meter reading.
This is where the basic problem lies: total consumption does not say where the cost comes from. It does not say which process stage wastes the most, when load should be shifted, or when consumption is out of line with the production level. Without process-level visibility, energy saving programmes easily remain general targets that never turn into anything concrete.
The requirements for energy management in industry are in a completely different class from the building side:
- Attribution by process: which device or stage generates the cost
- Context-specific interpretation: the same consumption can be normal or abnormal depending on load and conditions
- Real-time response: a cost deviation has to be spotted during the shift, not at the end of the month
- Explainability: the saving potential has to be demonstrable in euros to management and to customers
Without these capabilities energy management software is just a reporting tool — not a real basis for decision-making.
Five barriers that keep energy costs high
Many plants recognise the energy saving potential but cannot act on it. Here are the most common reasons.
1. The data is at too coarse a level
The monthly reading from the main electricity or fuel meter does not say which process stage the waste comes from. Without data per measurement point or per process stage, guessing is all that is left.
2. Data and production context live in different systems
Energy data sits in the energy meter or the automation system. Production volumes sit in the ERP or a separate spreadsheet. The two do not meet automatically — and combining them by hand happens with a month's delay at best.
3. Deviations are spotted too late
By the time an energy efficiency deviation shows up in a weekly report or on the monthly invoice, significant cost has already accumulated. Real-time monitoring is essential if the deviation is to be addressed during the same shift.
4. The analysis requires a specialist
In the traditional model energy data is exported to Excel or a BI tool, from which a specialist reports the findings. That is slow, expensive and dependent on one individual. When the specialist moves on, the knowledge goes with them.
5. The cloud is not an option
Cloud-based analysis sounds attractive, but in industry you run into security requirements, contractual obligations and continuity-of-supply requirements. Many plants cannot — or will not — move production data to an outside service.
How energy management software and local AI solve the problem
What works is not more reports or another dashboard. It is an architecture that brings data acquisition, contextualisation and the course of action into the same process.
OPC UA integration: data straight from the process
When energy management software connects directly to the automation system using the OPC UA protocol, energy data is available in real time without manual transfers. Every measurement point — power, flow, temperature, pressure, operating state — is written to the database with a timestamp, continuously.
Time-series storage: history available in seconds
An industrial time-series database is designed to store and query billions of rows. Data from years back can be retrieved in seconds. That is a precondition both for spotting trends and for demonstrating the saving potential in euros.
Local AI: context-specific savings analysis
A local language model analyses the energy data in its process context: what consumption is normal at this load and under these conditions, what the deviation most likely stems from, and how large the saving potential is in euros. The analysis happens on the plant's own network — the data does not leave.
An operational model: from decision to action
An observation alone is not enough. Dashboards, alarms and automated reports connect the observation to the right role at the right time. Maintenance gets the information during the shift, management gets a weekly review in euros, the customer gets an auditable report.
- 2–5 h
- saved each week on energy reporting when manual collection into Excel and formatting are replaced by automated reports. This is the typical change that energy management software brings right after commissioning.
The five-level model for finding saving potential
Looking at consumption trends alone does not tell you what to do. What is needed is an analysis model that moves from data to proposed actions. Here are the five levels:
Level 1: Baseline consumption per load level
Calculate the typical energy consumption per tonne, per hour or per batch of production in each load class. This is the reference point against which deviations are compared. Without a baseline there is no way to say whether consumption is normal or abnormal.
Level 2: Grouping by operating state
Segment the data by time of day, production recipe, day of the week and outdoor temperature. Night-time consumption differs from daytime, winter behaviour from summer. Without segmentation the model produces too many false findings.
Level 3: Correlation analysis
Which measurement points does the rise in energy cost relate to most? Did steam consumption rise without a matching increase in production? Did the temperature difference across the heat exchanger change? Correlation analysis shows where to look.
Level 4: Prioritisation in euros
Not all deviations matter equally. AI assesses the cost impact of the deviations and prioritises them in euros. That directs resources to where the saving potential is greatest.
Level 5: Closed feedback — from action to result
Once an action has been taken, its effect shows in the data. This closed loop is a precondition for continuous improvement — and for demonstrating to customers and management that the energy efficiency programme delivers results.
Practical example: a process industry cost increase nobody could explain
Consider a process industry plant where steam is produced by two boilers and heat is recovered in several heat exchangers between stages. Total consumption looks normal at monthly level, but the cost per tonne produced climbs steadily.
Process industry: from an unexplained cost increase to systematic energy management
Practical example
Before (coarse monitoring)
- Total consumption monitored — not at process level
- Cost per tonne up 3% in six months
- The cause was unknown — "maybe the raw material changed"
- An energy review once a month, done by a consultant
- Actions were generic: "use energy more efficiently"
- No way to show whether any action had made a difference
After (energy management software + AI)
- Steam consumption per tonne produced monitored in real time
- AI identifies repeated excess steam use in a particular load window
- Correlation analysis: the heat exchanger delta-T grows on the evening shift
- Likely cause: the control parameters do not respond optimally to changes
- Action: fine-tuning of the control parameters + updated monitoring
- Result: consumption levelled out, cost per tonne fell 2.8% in three weeks
- 2.8%
- drop in energy cost per unit of product in three weeks — a typical first result when process-level data and AI analysis are combined with a clear course of action. At smaller plants this means tens of thousands of euros a year.
On-premises versus cloud: which suits an industrial plant?
The cloud is not automatically a bad choice, but industry has typical constraints that make an on-premises implementation the only realistic option.
| Question | Cloud-based model | Local model (on-premises) |
|---|---|---|
| Where does the data live? | On the provider's servers | On the plant's own network |
| Security requirements | Requires a risk analysis and contracts | Data does not leave — easier to get approved |
| Operation during a network outage | Does not work without an external connection | Runs completely autonomously |
| Contractual requirements on data location | May conflict with customer contracts | Data stays in the agreed location |
| Cyber security policies | Production network data goes out — a risk | Production network stays isolated |
| Control over the infrastructure | Dependent on the provider | Own server, own control |
DataPortia™ implements this architecture in one package: an OPC UA interface, time-series storage in TimescaleDB and Ollama-based local AI analysis. Everything runs on the plant's own network — no cloud, no external services in normal operation.
How do you choose the right energy management software for industry?
If you have concluded that coarse monthly monitoring is no longer enough, the next step is choosing the right software. Here are five criteria:
- OPC UA support: make sure the software connects directly to the automation system with no intermediaries. This is a precondition for real-time data.
- Process-level granularity: the data must be available per measurement point, not just as a main meter reading. Otherwise locating deviations is impossible.
- Local AI: the analysis has to happen on the plant network. In many environments cloud analytics is neither possible nor permitted.
- Commissioning without an IT project: it must be possible to put the software into use quickly, without a months-long integration project.
- Explainability: the AI must produce explainable, auditable findings — not black boxes that cannot be justified to maintenance or to management.
A 90-day starting model: from energy savings to concrete results
Energy cost optimisation does not require a massive project. Here is a realistic way to start:
Month 1: Data basis and visibility
Pick one critical process area: steam production, compressed air or heat recovery, for example. Include 10–20 measurement points — power, flows, temperatures, operating states. Build the baseline consumption per load level. This gives you process-level visibility into energy consumption for the first time.
Month 2: Deviation detection and saving potential
Start deviation detection and AI analysis. The first findings typically come quickly — excess consumption in a particular load window, recurring waste heat or sub-optimal control. Prioritise them in euros.
Month 3: Actions and demonstrating the results
Carry out the first actions: fine-tuning the control parameters, updating monitoring thresholds or revising a maintenance date. Measure the effect — cost per unit of product before and after. This is the first clear ROI figure that can be put in front of management.
Summary: energy management software turns data into savings
In 2026 an industrial plant that tracks energy only at the level of the monthly invoice leaves significant saving potential on the table. Automation systems produce enormous amounts of energy data — the only question is whether it is collected and analysed properly.
Professional energy management software automates data acquisition, puts consumption in context by process state, identifies the saving potential with AI assistance and offers explainable proposals for action — all locally, with no cloud dependency.
Making the move does not mean a year-long IT project. It means choosing the right software, establishing the OPC UA connection, selecting 10–20 measurement points and building the baseline for the first process area. After that the saving potential starts to surface on its own.