Building anomaly

Ranks chilled-water operating patterns from meter, weather, and building proxies. No equipment telemetry or failure labels. Built by Alvin Alias.

497 units / 2.88M readings Isolation Forest with an LOF cross-check · no failure labels
AA-2026-04
model waking, first run may take up to a minute
Fleet snapshot · 497 units

The fleet · 497 units, scored from the meter study

pulling the fleet

-- high attention
-- elevated
-- watch
-- normal
-- total

Draws a complete historical reading from curated normal, watch, elevated, and high-attention examples.

Enter readings yourself 


        

          Recent pulls
          

            
Recent pulls

            

              
            

          

        

      


      

        

          --
          relative anomaly score
          
        


        


        

        
        

        

          What moved this score
          

            
What moved this reading score

            

          

        

      

    


    

      

        Calibration record · sensitivity checks and detector agreement
        

          
            
2,876,400 hourly readings scored · 4.18M raw, filtered for rolling-history coverage

            
497 units with at least 90 days of coverage

            
Anomaly rate 5.0% at contamination 0.05 · every point flagged at 0.02 stays flagged at 0.05

            
Unit baseline scores 35.0 to 74.0, median 58.6 · fleet strip shown as rank-relative bands

            
LOF vs Isolation Forest agreement 91.3% on a 100k sample

            
Top factor: 24-hour cooling-proxy variability; this explains detector score, not confirmed equipment faults

          
        

      

      

        How it works · building-meter and weather proxies
        
The detector combines chilled-water demand, outdoor weather, square footage, time, and rolling context. HVAC experience informed the hypotheses, but the source data do not contain input power, equipment temperatures, runtime, work orders, alarms, or failures.

      

      

        Notes · no failure labels exist; this reads anomaly structure
        
    
          
  1. This ML demo sleeps after extended inactivity. First wake can take a moment; runs after that are quick.
    2. 
          
  2. No failure labels exist in this data. The score reads anomaly structure, and the LOF cross-check is the only independent witness, which is why it prints with every run.
    3. 
          
  3. The fleet board is a scored snapshot of the study data, not a live feed; every unit on it is real and every score was computed by the model.
    4. 
          
  4. Demo scoring uses complete historical feature rows with rolling context; manual scoring estimates rolling context from the entered proxy values.
    5.