ML Load predictor, external temperature components by springfall2008 · Pull Request #3310 · springfall2008/batpred

springfall2008 · 2026-02-02T20:43:42Z

No description provided.

Copilot

Pull request overview

Adds an ML-based load forecasting feature (NumPy MLP) plus an external temperature forecast component, wiring both into Predbat’s forecasting pipeline and UI/docs.

Changes:

Introduces ML load predictor + load_ml component and publishes new HA sensors for forecast/stats.
Adds temperature component using Open-Meteo forecasts, and uses it as an ML feature and for charting.
Updates fetch pipeline, web charts, and mkdocs documentation/navigation.

Reviewed changes

Copilot reviewed 17 out of 21 changed files in this pull request and generated 27 comments.

Show a summary per file

File	Description
mkdocs.yml	Adds new ML load prediction page to docs nav
docs/load-ml.md	New ML load prediction documentation
docs/components.md	Documents new `temperature` and `load_ml` components
coverage/debug_predict.py	Debug helper for ML predictor behavior
coverage/debug_model.py	Debug helper for inspecting model weights/normalization
coverage/analyze_periods.py	Debug helper for analyzing dataset periods
coverage/analyze_data.py	Debug helper for basic dataset stats
apps/predbat/web.py	Adds LoadML charts and ML+PV+temperature visualization
apps/predbat/utils.py	Extends `prune_today()` for offset/future-window pruning
apps/predbat/unit_test.py	Registers new ML + temperature tests
apps/predbat/tests/test_temperature.py	New temperature component tests
apps/predbat/tests/test_minute_data_import_export.py	Updates call signature for `minute_data_import_export()`
apps/predbat/tests/test_load_ml.py	Comprehensive ML predictor + component tests
apps/predbat/temperature.py	New Open-Meteo temperature forecast component
apps/predbat/predbat.py	Version bump; adds `temperature.py` to `PREDBAT_FILES`
apps/predbat/load_predictor.py	New NumPy-only MLP predictor (train/predict/persist)
apps/predbat/load_ml_component.py	New component wrapper for data fetch/train/publish predictions
apps/predbat/fetch.py	Adds ML forecast integration; changes `minute_data_import_export()` signature
apps/predbat/config.py	Adds `load_ml_enable` schema entry
apps/predbat/components.py	Registers new `temperature` and `load_ml` components
.cspell/custom-dictionary-workspace.txt	Adds ML-related words for spellchecker

Copilot · 2026-02-07T17:22:25Z

apps/predbat/load_ml_component.py

+        except Exception as e:
+            self.log("Error: ML Component: Failed to fetch load data: {}".format(e))
+            print("Error: ML Component: Failed to fetch load data: {}".format(e))
+            import traceback
+
+            self.log("Error: ML Component: {}".format(traceback.format_exc()))
+            return None, 0, 0, None, None


This exception handler logs via self.log(...) but also prints directly to stdout. Predbat uses self.log() for logging; print() can bypass HA/AppDaemon logging and is easy to miss. Remove the print() and rely on self.log() (the traceback is already logged).

Copilot · 2026-02-07T17:22:26Z

apps/predbat/load_predictor.py

+                batch_weights = weights_shuffled[batch_start:batch_end]
+
+                # Forward pass
+                y_pred, activations, pre_activations = self._forward(X_batch)
+
+                # Apply sample weights to loss (approximate by weighting gradient)
+                weighted_y_batch = y_batch * batch_weights.reshape(-1, 1)
+                weighted_y_pred = y_pred * batch_weights.reshape(-1, 1)
+
+                batch_loss = mse_loss(y_batch, y_pred)
+                epoch_loss += batch_loss
+                num_batches += 1
+
+                # Backward pass
+                weight_grads, bias_grads = self._backward(y_batch, activations, pre_activations)


Sample weights are computed/shuffled (train_weights, batch_weights) and even used to create weighted_y_batch/weighted_y_pred, but those weighted tensors are never used in the loss or gradient calculation. As a result, the intended time-decay weighting has no effect. Either incorporate batch_weights into the loss/gradient (e.g., weight the per-sample loss or scale delta), or remove the unused weighting code to avoid a false sense of correctness.

Copilot · 2026-02-07T17:22:26Z

apps/predbat/load_predictor.py

+        # Validation uses most recent data (minute 0 to validation_holdout)
+        # Training uses ALL data (minute 0 to end_minute), including validation period
+        validation_end = validation_holdout_hours * 60


The dataset builder explicitly includes the validation period in the training set (Training uses ALL data ... including validation period). This makes the reported validation MAE overly optimistic and can cause the component to accept models that don’t generalize. Hold out the validation window from training (or rename this metric to something like “fit_mae”) so the gating threshold reflects true out-of-sample performance.

Copilot · 2026-02-07T17:22:26Z

apps/predbat/utils.py

@@ -54,18 +54,19 @@ def prune_today(data, now_utc, midnight_utc, prune=True, group=15, prune_future=
            timekey = datetime.strptime(key, TIME_FORMAT_SECONDS)
        else:
            timekey = datetime.strptime(key, TIME_FORMAT)
-        if last_time and (timekey - last_time).seconds < group * 60:
+        if last_time and (timekey - last_time).total_seconds() < group * 60:
            continue
-        if intermediate and last_time and ((timekey - last_time).seconds > group * 60):
+        if intermediate and last_time and ((timekey - last_time).total_seconds() > group * 60):
            # Large gap, introduce intermediate data point
            seconds_gap = int((timekey - last_time).total_seconds())
            for i in range(1, seconds_gap // int(group * 60)):
-                new_time = last_time + timedelta(seconds=i * group * 60)
-                results[new_time.strftime(TIME_FORMAT)] = prev_value
+                new_time = last_time + timedelta(seconds=i * group * 60) + timedelta(minutes=offset_minutes)
+                results[new_time.isoformat()] = prev_value
        if not prune or (timekey > midnight_utc):
-            if prune_future and (timekey > now_utc):
+            if prune_future and (timekey > (now_utc + timedelta(days=prune_future_days))):
                continue
-            results[key] = data[key]
+            new_time = timekey + timedelta(minutes=offset_minutes)
+            results[new_time.isoformat()] = data[key]
            last_time = timekey


prune_today() now rewrites all timestamp keys using datetime.isoformat(), which changes key formatting and breaks callers/tests that rely on keys being preserved (e.g., apps/predbat/tests/test_prune_today.py asserts original keys are present). Consider preserving the original key strings when no shifting is requested, and when shifting/creating intermediate points, format new keys using the existing TIME_FORMAT / TIME_FORMAT_SECONDS conventions rather than isoformat() to avoid inconsistent timestamp formats across the app.

Copilot · 2026-02-07T17:22:26Z

apps/predbat/fetch.py

+        if self.get_arg("load_ml_enable", False) and self.get_arg("load_ml_source", False):
+            load_ml_forecast = self.fetch_ml_load_forecast(self.now_utc)
+            self.load_forecast_only = True  # Use only ML forecast for load if enabled


fetch_sensor_data() sets self.load_forecast_only = True whenever ML is enabled/configured, even if fetch_ml_load_forecast() returns an empty dict (e.g., ML entity missing/unavailable). That will skip historical filtering/adjustment paths gated by not self.load_forecast_only, potentially degrading planning. Only set load_forecast_only when a non-empty ML forecast was successfully loaded.

Suggested change

if self.get_arg("load_ml_enable", False) and self.get_arg("load_ml_source", False):

load_ml_forecast = self.fetch_ml_load_forecast(self.now_utc)

self.load_forecast_only = True # Use only ML forecast for load if enabled

# Default to using historical-based load forecast unless a valid ML forecast is loaded

self.load_forecast_only = False

if self.get_arg("load_ml_enable", False) and self.get_arg("load_ml_source", False):

load_ml_forecast = self.fetch_ml_load_forecast(self.now_utc)

# Only rely solely on ML forecast if we actually received one

if load_ml_forecast:

self.load_forecast_only = True # Use only ML forecast for load if enabled and available

Copilot · 2026-02-07T17:22:30Z

apps/predbat/tests/test_load_ml.py

+    Test training on real load_minutes_debug.json data and generate comparison chart
+    """
+    import json
+    import os


This import of module os is redundant, as it was previously imported on line 15.

Suggested change

import os

Copilot · 2026-02-07T17:22:30Z

apps/predbat/tests/test_temperature.py

+from datetime import datetime, timezone
+
+
+class MockTemperatureAPI(TemperatureAPI):


This class does not call ComponentBase.init during initialization. (MockTemperatureAPI.init may be missing a call to a base class init)

Copilot · 2026-02-07T17:22:30Z

apps/predbat/tests/test_temperature.py

+            temperature_url=temperature_url
+        )
+
+    def log(self, message):


This method is shadowed by attribute log in superclass ComponentBase.

Copilot · 2026-02-07T17:22:31Z

apps/predbat/tests/test_load_ml.py

+                plt.savefig(chart_path, dpi=150, bbox_inches="tight")
+                print(f"  Chart saved to {chart_path}")
+                break
+            except:


Except block directly handles BaseException.

Suggested change

except:

except Exception:

Copilot · 2026-02-07T17:22:31Z

apps/predbat/load_ml_component.py

+                    self.load_minutes_now = load_minutes_now
+                    self.data_ready = True
+                    self.last_data_fetch = self.now_utc
+                    pv_data = pv_data


This assignment assigns a variable to itself.

Suggested change

pv_data = pv_data

Web status update more rapid

springfall2008 added 19 commits January 30, 2026 08:11

WIP

5e7b908

WIP

c428c8b

Merge branch 'aiload' of github.com:springfall2008/batpred into aiload

f73f54f

WIP

e4afb22

Merge branch 'aiload' of github.com:springfall2008/batpred into aiload

00bd7cc

WIP

b61a5ac

Fix car subtract issues

5feaee3

Fixes for load

9da7676

Chart fixes

07f4343

chart fixes

ba53b2d

Chart fixes

620edbc

Version

0307094

Schema fix

85ad484

Adding PV to model

5bf8f4a

Temperature API

6844a4b

Hook temperature into ML inputs

85fdf1e

Fixes for temp corrolation

64bd47e

Docs

6cefa95

Chart tidy

756b5a2

springfall2008 marked this pull request as ready for review February 7, 2026 17:15

Copilot AI review requested due to automatic review settings February 7, 2026 17:15

Copilot started reviewing on behalf of springfall2008 February 7, 2026 17:16 View session

Copilot AI reviewed Feb 7, 2026

View reviewed changes

springfall2008 added 2 commits February 7, 2026 17:47

Charting power for load ml

ddf836f

Dead code

e4ed1f1

springfall2008 changed the title ~~ML Load predictor~~ ML Load predictor, external temperature components Feb 7, 2026

Review feedback

05f3748

Web status update more rapid

springfall2008 merged commit 1972960 into main Feb 7, 2026
1 check passed

springfall2008 deleted the aiload branch February 7, 2026 18:01

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

ML Load predictor, external temperature components#3310

ML Load predictor, external temperature components#3310
springfall2008 merged 22 commits intomainfrom
aiload

springfall2008 commented Feb 2, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Feb 7, 2026

Uh oh!

Copilot AI Feb 7, 2026

Uh oh!

Copilot AI Feb 7, 2026

Uh oh!

Copilot AI Feb 7, 2026

Uh oh!

Copilot AI Feb 7, 2026

Uh oh!

Copilot AI Feb 7, 2026

Uh oh!

Copilot AI Feb 7, 2026

Uh oh!

Copilot AI Feb 7, 2026

Uh oh!

Copilot AI Feb 7, 2026

Uh oh!

Copilot AI Feb 7, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

		from datetime import datetime, timezone


		class MockTemperatureAPI(TemperatureAPI):

Uh oh!

Conversation

springfall2008 commented Feb 2, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Feb 7, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 7, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 7, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 7, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 7, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 7, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 7, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 7, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 7, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 7, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant