diff --git a/india_forecast_app/models/pvnet/model.py b/india_forecast_app/models/pvnet/model.py
index 6a1e38a..fff901e 100644
--- a/india_forecast_app/models/pvnet/model.py
+++ b/india_forecast_app/models/pvnet/model.py
@@ -11,7 +11,7 @@
 import numpy as np
 import pandas as pd
 import torch
-from ocf_datapipes.batch import batch_to_tensor, copy_batch_to_device, stack_np_examples_into_batch
+from ocf_datapipes.batch import batch_to_tensor, copy_batch_to_device, stack_np_examples_into_batch, BatchKey
 from ocf_datapipes.training.pvnet_site import construct_sliced_data_pipeline as pv_base_pipeline
 from ocf_datapipes.training.windnet import DictDatasetIterDataPipe, split_dataset_dict_dp
 from ocf_datapipes.training.windnet import construct_sliced_data_pipeline as wind_base_pipeline
@@ -95,6 +95,10 @@ def predict(self, site_id: str, timestamp: dt.datetime):
             for i, batch in enumerate(self.dataloader):
                 log.info(f"Predicting for batch: {i}")
 
+                if self.name == 'windnet_ad_sites_generation_delay':
+                    # this is a bit of hack, but it's important to do what was done in training
+                    batch[BatchKey.wind][int(batch[:, BatchKey.wind_t0_idx])] = -1
+
                 # save batch
                 save_batch(batch=batch, i=i, model_name=self.name, site_uuid=self.site_uuid)