Projects / Time series prediction
Pump efficiency and water flow predictions
Water flows, leaks and pump efficiencies.
Keywords
Machine learning
Water industry
Digitalization
Predictive analysis
Time series analysis
Evaluation
Introduction
VASYD manages water and waste water in Southern Skåne, Sweden. VASYD is one of Sweden's largest water and waste management organizations that invests heavily in sustainable sewage treatment and pushes for environmentally smart solutions in community building.
VASYD's mission is to purify water, ensure water quality and deliver drinking water while taking care of what is flushed down our drains in order to care for oceans and waterways.
Challenge
As much as 20% of drinking water is lost due to leakage. With data sets from different geographical locations in multiple formats relating to several aspects of water and water operations, VASYD wanted to evaluate the potential in existing data to combat this problem. How can we become better by using what we have? Relevant topics where today's solutions are outdated or insufficient included:
Flow predictions (how much water will be needed in area X at time T?)
Anomaly detection (how can we detect when something goes wrong?)
Pump efficiency (how can water pumps alternate smarter to reduce efficiency loss?)
Goal
Use provided data sets and evaluate the potential in data by leveraging the power of statistical analysis and machine learning.
Solution
We gathered and analyzed the provided datasets with historical data on water flow rates, pump usage, and other relevant variables. We then used statistical analysis (primarily time series analysis) and machine learning algorithms to identify patterns and trends in the data. This allowed us to develop accurate predictions of water flow, as well as identify potential anomalies and areas for optimization.
Statistical time series modelling. Green line is predicted future water flow. Black dots are actual data.
Next, we worked with water industry experts to identify the most significant areas for improvement and to understand the operations aspect and find the potential improvements. We then tested these solutions using real-world data to verify their accuracy and effectiveness.
We found that predictive maintenance for water pumps is possible with statistical analysis and machine learning. By using predictive algorithms, we can forecast when maintenance is needed, thereby reducing downtime and extending the life of equipment.
Furthermore, by implementing algorithms to rotate water pump usage, we can reduce wear and improve the overall efficiency of the system. Finally, water flow predictions could be fed to the water pumps to give pump control system AI capabilities. This would enable the system to make real-time adjustments based on demand and other factors, leading to even greater efficiency gains.
Visualization of water pump efficiency over time. Red, blue, green are different pumps that rotate to keep a certain water flow. Pumps wear out over time - if they had access to more information about incoming water flows they could rotate better to reduce power required to the the same throughput.
Results
Successful statistical analysis and machine learning proof of concepts concluded five different focus areas:
Water flow predictions are not just possible, but very accurate
Dynamic alarm levels for anomalies can be implemented
Predictive maintenance for water pumps is possible
Algorithms to rotate water pump usage would reduce wear
Water flow predictions could be fed to the water pumps to give pump control system AI capabilities.
Our analysis became part of the report Digitizing the Swedish water industry where the potential for machine learning using existing water data is described in more detail.