v8.33.3, ML: in LoadML Power chart, little variation in LoadMLpower with temperature

[SwiftRR]

Describe the bug
In the LoadMLPower chart, the predicted power hardly changes with changing temperature. I am sure in earlier versions that the predicted power decreased with increasing temperature power (and vice versa).
I'm certainly seeing some variation in my ML power chart with varying temperature, but not a massive variation

Expected behaviour
A larger variation of load power with temperature.
For me, with an ASHP, temperature is an enormous factor in determining the amount of energy/power

Predbat version

v8.33.3

Environment details

• GE AC3.0 with two 8.2 kWh batteries
• Standard HAOS Core 2026.2.2; OS 17.1, running on Raspberry Pi

Screenshots
If applicable, add screenshots to help explain your problem. The most useful ones can be your battery chart, the Predbat HTML plan and your current settings in HA.

Log file
Can you capture a log file from the time of the issue, debug mode is not normally required.

apps.yaml.txt

predbat-17.log

predbat_plan.html.txt

predbat_debug.yaml.txt

Rob

[springfall2008]

What does the history show, can you see temperature changes impacting load going back 7 days?

[SwiftRR]

This is my LoadMLPower chart on v8.33.3 with history back to 13th Feb.
I am on IOG and you can see the overnight cheap slots when the batteries were charging.

There seems to be little variation in ML predicted future for Load Power with changes in estimated future temperatures.

I updated to v8.33.3 from v8.33.1 on 17th Feb at 21:11.

This is a screenshot from v8.33.1. The ML predicted future for Load Power shows a marked decrease with increasing temperature power (and vice versa). Something that the v8.33.3 chart does not show.

So has something changed here between v8.33.1 and v8.33.3 (Sorry, I missed v8.33.2)

Rob

[Sycamore967]

I only enabled ML 3 days ago and I'm not using it for predictions yet. However, it seems to be stable and showing me what I'd expect. I'm posting as I'm also on IOG. I don't have any heat load, so I wouldn't expect a massive load change with temp. The biggest difference for me would be less heating pump use as temperature increases and its not on so much.

v8.33.3

[SwiftRR]

I can add that I do have an ASHP and its load is greatly affected by temperature. So I would expect that future load predicted would be lower with increasing temperature – indeed this would be one of the advantages of ML for future predictions of energy use.

Rob

[springfall2008]

You might want to try the new version on 'main' that I'm testing

• Uses up to 28 days history for training
• Improved incremental training

[SwiftRR]

I have tried the version on 'main' but I completely lose Load (ML Forecast) 'red' lines in the LoadML chart.
In the LoadMLPower chart, there is no trace for Load Power (Predicted future).
Predbat.log shows that ML has not run at all with this version.
In Dash, status showed as 'demand (stale)'
I left the version running for an hour but no difference.

I have now downgraded back to v8.33.3 and predbat.log shows that the ML model is once again running.
The charts, the red lines are back in the two ML charts..

So an unsuccessful experiment, I am afraid.

Rob

[springfall2008]

New version on main worth trying again, it maybe you are missing history for pv_power or load_power?

[SwiftRR]

I have updated to main from v8.33.3
I do now have predicted lines showing but there is still very little correlation between temperature and Load power future.

Charts

Predbat.log shows that ML is training.
I will leave this version running overnight.

2026-02-21 17:34:36.306033: ML Component: Fetching 28 days of load history from ['sensor.givtcp_ce2218g352_load_energy_today_kwh']
2026-02-21 17:34:36.873396: Processing 1344 30-minute periods for power integration
2026-02-21 17:34:36.931439: ML Component: Applying improved car charging subtraction with hold and interpolation
2026-02-21 17:34:37.174440: ML Component: Car charging subtraction applied, 8064 data points after subtraction
2026-02-21 17:34:37.352559: ML Temperature data points: 19052
2026-02-21 17:34:37.430673: ML Component: Fetched 8064 load data points, 7.0 days of history
2026-02-21 17:34:37.434113: ML Component: Starting fine-tune training (2h interval)
2026-02-21 17:34:37.434197: ML Predictor: Starting fine-tune training with 20 epochs
2026-02-21 17:34:37.449435: ML Predictor: Creating dataset with 671 hours of training samples and 48 hours of validation samples
2026-02-21 17:34:40.064358: ML Predictor: Created 7774 training samples, 576 validation samples
2026-02-21 17:34:40.176141: ML Predictor: Reset Adam optimizer state for fine-tuning
2026-02-21 17:34:40.220089: ML Predictor: Baseline (pre-finetune) val_mae=0.0172 kWh
2026-02-21 17:34:44.273712: ML Predictor: Epoch 1/20: train_loss=1.3942 val_mae=0.0139 kWh
2026-02-21 17:34:48.116569: ML Predictor: Epoch 2/20: train_loss=0.3504 val_mae=0.0132 kWh
2026-02-21 17:34:51.858445: ML Predictor: Epoch 3/20: train_loss=0.2219 val_mae=0.0098 kWh
2026-02-21 17:34:55.509744: ML Predictor: Epoch 4/20: train_loss=0.1919 val_mae=0.0103 kWh
2026-02-21 17:34:59.168915: ML Predictor: Epoch 5/20: train_loss=0.1573 val_mae=0.0140 kWh
2026-02-21 17:35:02.812683: ML Predictor: Epoch 6/20: train_loss=0.1906 val_mae=0.0100 kWh
2026-02-21 17:35:06.571191: ML Predictor: Epoch 7/20: train_loss=0.1473 val_mae=0.0086 kWh
2026-02-21 17:35:10.530315: ML Predictor: Epoch 8/20: train_loss=0.1262 val_mae=0.0093 kWh
2026-02-21 17:35:14.201530: ML Predictor: Epoch 9/20: train_loss=0.0932 val_mae=0.0084 kWh
2026-02-21 17:35:17.815776: ML Predictor: Epoch 10/20: train_loss=0.1143 val_mae=0.0089 kWh
2026-02-21 17:35:21.804109: ML Predictor: Epoch 11/20: train_loss=0.1278 val_mae=0.0073 kWh
2026-02-21 17:35:25.468700: ML Predictor: Epoch 12/20: train_loss=0.0918 val_mae=0.0063 kWh
2026-02-21 17:35:29.076790: ML Predictor: Epoch 13/20: train_loss=0.0808 val_mae=0.0084 kWh
2026-02-21 17:35:32.660581: ML Predictor: Epoch 14/20: train_loss=0.1019 val_mae=0.0113 kWh
2026-02-21 17:35:36.349933: ML Predictor: Epoch 15/20: train_loss=0.0831 val_mae=0.0068 kWh
2026-02-21 17:35:40.005453: ML Predictor: Epoch 16/20: train_loss=0.0705 val_mae=0.0063 kWh
2026-02-21 17:35:43.601901: ML Predictor: Epoch 17/20: train_loss=0.0617 val_mae=0.0051 kWh
2026-02-21 17:35:47.267407: ML Predictor: Epoch 18/20: train_loss=0.0929 val_mae=0.0060 kWh
2026-02-21 17:35:50.878375: ML Predictor: Epoch 19/20: train_loss=0.0658 val_mae=0.0060 kWh
2026-02-21 17:35:54.485309: ML Predictor: Epoch 20/20: train_loss=0.0656 val_mae=0.0056 kWh
2026-02-21 17:35:54.485392: ML Predictor: Restored best weights from epoch with val_mae=0.0051 kWh
2026-02-21 17:35:54.485413: ML Predictor: Training complete, final val_mae=0.0051 kWh
2026-02-21 17:35:54.510933: ML Component: Training successful, val_mae=0.0051 kWh
2026-02-21 17:35:54.558158: ML Predictor: Model saved to /config/predbat_ml_model.npz
2026-02-21 17:35:55.236293: ML Component: Generated 576 predictions (total 39.96 kWh over 48h)
2026-02-21 17:35:55.255319: ML Component: Prediction cycle completed

predbat.log.txt
predbat_debug.yaml.txt

Rob