15 KiB
Tasks
-
Quantiles zelf breder maken na fitten, literatuur bekijken (overconfident voor ondergrens) (Validation set januari 2023) Eerst literatuur bekijken ofdat probleem al voorkomt
-
compare reconstructed prices with real imbalance prices (on figure and metrics on whole test set)
-
Baseline policy (Eche imbalanceprijzen):
- Batterij:
- 2 MWh
- 1 MW power (charging / discharging)
- Training data -> 2 tresholds (1 voor charging, 1 voor discharging) (Volledig chargen / dischargen)
- 2 Mwh -> 8 kwartieren
2 MWh -> 0.25 MW per kwartier -> charge cycle: 4 MWh -> 16 kwartieren iets doen -> dan charge cycle (max 400 per jaar voor later)
- Batterij:
-
1000 generaties voor 1 dag uit test set -> 1000 prijzen reconstrueren -> tresholds bepalen (logische tresholds) -> 2 tresholds voor 1 dag en toepassen op de echte prijzen van de dag
-
Learning rate non autoregressive
-
!!! Historgram quantile plot volledige dag
-
CRPS en MAE, over 96 wanneer wordt het slecht
-
crps loss fixen
-
More input parameters
-
Non autoregressive (baseline for linear)
(
-
NRV to prices (bid ladder from yesterday) -> Elia
-
Simple policy )
-
Diffusion vs Variable Selection Network
-
Bid ladder van gisteren
-
Plot van reconstucted prices with real bid ladder + bid ladder from yersterday against real reconstructed prices
Presentatie:
- Korte motivatie -> op dit moment: veel batterijen op simpele manier zoals onze baseline
- Policy per dag afhankelijk van generation model (policy word day ahead vastgelegd)
- NRV -> resultaten van verschillende resultaten (verder werken met autoregresive om te kunnen samplen)
-> ook overconfidence tonen van de quantiles (met die histogram)
- Quantile regreession + figuur van CDF
Resultaten van de autoregression modellen
- Plot van CRPS (kijk teams)
- Prijs reconstructie + momenteel bid ladders geplubliceerd door Elia momenteel. (paar plotjes)
- Baseline Policy voor batterij
- (Resultaten van nieuwe methode met samplen)
- Volgende stappen:
3 mogelijkheden:
- bid ladders (niet echt doen)
- Generaties beter van NRV -> quantile regression + overconfidence fixen (+ betere modellen)
Diffusion model is er een van en niet probleem van autoregressive
- Betere RL modellen (policies)
(mag 20 minuten zijn)
Different Model Architectures (For Quantile Regression)
Baseline
The baseline just calculates the values for the given quantiles using the training data.
Mean NRV for whole day
| train_CRPSLoss | test_CRPSLoss |
|---|---|
| 70.70666326392862 | 76.03928650729421 |
Auto Regressive Models
Linear Model
Example summary of the Linear Model
==========================================================================================
Layer (type:depth-idx) Output Shape Param #
==========================================================================================
Sequential [1024, 13] --
├─TimeEmbedding: 1-1 [1024, 195] --
│ └─Embedding: 2-1 [1024, 2] 192
├─LinearRegression: 1-2 [1024, 13] --
│ └─Linear: 2-2 [1024, 13] 2,548
==========================================================================================
Total params: 2,740
Trainable params: 2,740
Non-trainable params: 0
Total mult-adds (M): 2.81
==========================================================================================
Input size (MB): 0.79
Forward/backward pass size (MB): 0.12
Params size (MB): 0.01
Estimated Total Size (MB): 0.93
==========================================================================================
| Experiment | Quarter | Day Of Week | Load | Wind | Net | test_L1Loss | test_CRPSLoss |
|---|---|---|---|---|---|---|---|
| Link | True | True | True | True | True | 101.63938813784577 | 68.4846434120529 |
| Link | True | True | True | False | False | 105.50163727880451 | 73.50646092092921 |
Quantiles Histogram Debug Plots
Non Linear Model
Example summary of the Non Linear Model
==========================================================================================
Layer (type:depth-idx) Output Shape Param #
==========================================================================================
Sequential [1024, 13] --
├─TimeEmbedding: 1-1 [1024, 96] --
├─NonLinearRegression: 1-2 [1024, 13] --
│ └─ModuleList: 2-9 -- (recursive)
│ │ └─Linear: 3-1 [1024, 512] 49,664
│ └─ReLU: 2-2 [1024, 512] --
│ └─ModuleList: 2-9 -- (recursive)
│ │ └─Dropout: 3-2 [1024, 512] --
│ └─ReLU: 2-4 [1024, 512] --
│ └─ModuleList: 2-9 -- (recursive)
│ │ └─Linear: 3-3 [1024, 512] 262,656
│ └─ReLU: 2-6 [1024, 512] --
│ └─ModuleList: 2-9 -- (recursive)
│ │ └─Dropout: 3-4 [1024, 512] --
│ └─ReLU: 2-8 [1024, 512] --
│ └─ModuleList: 2-9 -- (recursive)
│ │ └─Linear: 3-5 [1024, 13] 6,669
==========================================================================================
Total params: 318,989
Trainable params: 318,989
Non-trainable params: 0
Total mult-adds (M): 326.64
==========================================================================================
Input size (MB): 0.39
Forward/backward pass size (MB): 8.50
Params size (MB): 1.28
Estimated Total Size (MB): 10.16
==========================================================================================
| Experiment | Quarter | Day Of Week | Load | Wind | Net | test_L1Loss | test_CRPSLoss |
|---|---|---|---|---|---|---|---|
| Link | True | True | False | False | False | 103.6061492655297 | 67.56606704206783 |
| Link | True | True | True | False | False | 103.6867771851146 | 68.6250203464861 |
| Link | True | True | True | True | True | 102.03053663410438 | 68.968302991015 |
Quantiles Histogram Debug Plots
LSTM/GRU Model
Input shape: (batch_size, sequence_length, input_features)
If load forecast is used as input, each time step gets the load forecast for the next time step as feature.
Example: We have time steps:
| 1 | 2 | 3 | 4 | 5 | |
|---|---|---|---|---|---|
| NRV | 0.1 | 0.2 | 0.15 | 0.3 | 0.4 |
| Load forecast | 0.4 | 0.23 | 0.48 | 0.2 | 0.1 |
If we want to predict the NRV for time step 5 using the information we have beforehand, we can use the NRV from the previous time steps. We can however also use the load forecast of time step 5. To incorporate this information as input, we need to move the load forecast one time step back. This means, that the input for time step 5 is given with the NRV of time step 4.
If the time is also wanted as input, we add this as a feature for every timestep aswell.
Example summary of the LSTM/GRU Model
==========================================================================================
Layer (type:depth-idx) Output Shape Param #
==========================================================================================
Sequential [512, 13] --
├─TimeEmbedding: 1-1 [512, 96, 5] --
│ └─Embedding: 2-1 [512, 96, 4] 384
├─GRUModel: 1-2 [512, 13] --
│ └─GRU: 2-2 [512, 96, 512] 3,949,056
│ └─Linear: 2-3 [512, 13] 6,669
==========================================================================================
Total params: 3,956,109
Trainable params: 3,956,109
Non-trainable params: 0
Total mult-adds (G): 194.11
==========================================================================================
Input size (MB): 0.39
Forward/backward pass size (MB): 202.95
Params size (MB): 15.82
Estimated Total Size (MB): 219.17
==========================================================================================
| Experiment | Quarter | DoW | Load | Wind | Net | test_L1Loss | test_CRPSLoss |
|---|---|---|---|---|---|---|---|
| Link | True | True | True | True | True | 104.26072751090527 | 66.05181515234297 |
Quantiles Histogram Debug Plots
Results for now
| Model | test_L1Loss | test_CRPSLoss |
|---|---|---|
| Linear Model | 101.63938813784577 | 68.4846434120529 |
| Non Linear Model | 102.03053663410438 | 68.968302991015 |
| LSTM/GRU Model | 104.26072751090527 | 66.05181515234297 |
Non Auto-Regressive Models
| Model | Experiment | test_L1Loss | test_CRPSLoss |
|---|---|---|---|
| Linear Model | Link | 104.89022124436754 | 69.04472427024562 |
| Non Linear Model | Link | 103.3332725941881 | 69.50204645931149 |