Teburin Abubuwan Ciki
- 1. Gabatarwa
- 2. Hanyoyin Aiki
- 3. Gwaje-gwaje da Sakamako
- 4. Cikakkun Bayanai na Fasaha
- 5. Misalin Tsarin Bincike
- 6. Aikace-aikace na Gaba da Hasashe
- 7. Sharhin Masana
- 8. Bincike na Asali
- 9. Manazarta
1. Gabatarwa
Ƙarfin samar da wutar lantarki ta hanyar Photovoltaic (PV) yana canzawa sosai saboda dogaro da yanayin yanayi kamar rufe gajimare, matsayin rana, da iskar gas. Hasashen ɗan gajeren lokaci, ko hasashen nan take, a ma'aunin minti yana da mahimmanci ga sarrafa grid mai wayo, tabbatar da ci gaba da wutar lantarki, da sarrafa ƙimar hawan wutar. Hanyoyin gargajiya da suka dogara da hasashen yanayi na lambobi (NWP) ko bayanan tauraron dan adam galibi suna fama da ƙarancin ƙuduri na sarari da lokaci, wanda ya sa ba su dace da hasashen gida, na matakin minti ba. Wannan aikin yana ba da shawarar hanyar koyon zurfi wacce ke amfani da hotunan sama da aka ɗauka ta kyamara ta ƙasa da bayanan wutar PV na tarihi don hasashen ƙarfin wutar nan gaba.
2. Hanyoyin Aiki
2.1 Tsarin Matsala
Manufar ita ce hasashen ƙarfin wutar PV $P_{t+\Delta t}$ a wani lokaci na gaba $t+\Delta t$ (misali, minti 1 gaba) bisa ga ƙimar wutar tarihi $\{P_{t}, P_{t-1}, ..., P_{t-N}\}$ da jerin hotunan sama $\{I_{t}, I_{t-1}, ..., I_{t-M}\}$. Tsawon hasashen $\Delta t$ yawanci gajere ne (minti 1-10).
2.2 Tarin Bayanai da Shirye-shirye
An tattara bayanan a Kyoto, Japan, kuma ya ƙunshi ma'aunin wutar PV da hotunan sama masu dacewa da aka ɗauka a tazarar minti 1. Ana shirya hotuna ta hanyar yanke su zuwa wani yanki mai mahimmanci (ROI) a kusa da rana, sake girman su zuwa pixels 224x224, da daidaita ƙimar pixel. Ana daidaita ƙimar wutar zuwa kewayon [0,1] ta amfani da ma'aunin min-max.
2.3 Tsarin Sadarwa
Ana kwatanta tsarin koyon zurfi guda uku:
- MLP (Multi-Layer Perceptron): Samfurin tushe wanda ke ɗaukar ƙimar wutar tarihi kawai a matsayin shigarwa. Ya ƙunshi yadudduka 3 masu haɗin kai gabaɗaya tare da neurons 128, 64, da 1, bi da bi, ta amfani da kunnawa ReLU.
- CNN (Convolutional Neural Network): Yana sarrafa hotunan sama ta amfani da yadudduka 3 na convolution (fitarwa 32, 64, 128, girman kernel 3x3) sannan da max-pooling da yadudduka 2 masu haɗin kai gabaɗaya. Ana haɗa ƙimar wutar tarihi kafin ƙarshen Layer.
- LSTM (Long Short-Term Memory): Yana amfani da Layer LSTM tare da raka'a 128 na ɓoye don yin samfurin dogaro na lokaci a cikin jerin ƙimar wutar da fasalulluka na hoto da CNN ta fitar. Ana wucewa da fitar LSTM ta hanyar Layer mai haɗin kai gabaɗaya don samar da hasashen.
3. Gwaje-gwaje da Sakamako
3.1 Ma'aunin Kimantawa
Ana kimanta aiki ta amfani da Tushen Matsakaicin Kuskuren Murabba'i (RMSE) da ƙimar fasaha ta RMSE, wanda aka ayyana kamar:
$\text{Ƙimar Fasaha} = 1 - \frac{\text{RMSE}_{model}}{\text{RMSE}_{ciwawa}}$
inda tushen ciwawa ya ɗauka $P_{t+\Delta t} = P_t$.
3.2 Sakamakon Ƙididdiga
Teburin da ke ƙasa ya taƙaita ƙimar fasaha ta RMSE don hasashen minti 1 gaba:
| Samfuri | Ƙimar Fasaha ta RMSE (%) |
|---|---|
| Tushen Ciwawa | 0 |
| MLP | 7 |
| CNN | 12 |
| LSTM | 21 |
Samfurin LSTM ya sami mafi girman ƙimar fasaha na 21%, wanda ya fi na MLP da CNN sosai. Wannan yana nuna mahimmancin yin samfurin dogaro na lokaci don hasashen nan take.
3.3 Nazarin Cirewa
Gwaje-gwajen cirewa sun nuna cewa amfani da duka ƙimar wutar tarihi da hotunan sama yana ba da sakamako mafi kyau fiye da amfani da kowane ɗayan kaɗai. Samfurin LSTM kuma yana amfana daga dogon jerin shigarwa (har zuwa minti 10 na tarihi).
4. Cikakkun Bayanai na Fasaha
4.1 Tsarin Lissafi
Kwayar LSTM tana ƙididdige yanayin ɓoye $h_t$ da yanayin tantanin halitta $c_t$ kamar:
$f_t = \sigma(W_f \cdot [h_{t-1}, x_t] + b_f)$
$i_t = \sigma(W_i \cdot [h_{t-1}, x_t] + b_i)$
$\tilde{c}_t = \tanh(W_c \cdot [h_{t-1}, x_t] + b_c)$
$c_t = f_t \odot c_{t-1} + i_t \odot \tilde{c}_t$
$o_t = \sigma(W_o \cdot [h_{t-1}, x_t] + b_o)$
$h_t = o_t \odot \tanh(c_t)$
inda $x_t$ shine shigarwa a lokaci $t$, $\sigma$ shine aikin sigmoid, kuma $\odot$ yana nuna ninka ta kashi.
4.2 Aikin Asara da Ingantawa
Ana horar da samfuran don rage girman asarar Matsakaicin Kuskuren Murabba'i (MSE):
$\mathcal{L} = \frac{1}{N} \sum_{i=1}^{N} (P_i - \hat{P}_i)^2$
Ana yin ingantawa ta amfani da mai inganta Adam tare da ƙimar koyo na $10^{-4}$ da girman rukuni na 32. Ana amfani da dakatarwa da wuri bisa ga asarar tabbatarwa.
5. Misalin Tsarin Bincike
Yi la'akari da yanayin da manajan gonar hasken rana ke son hasashen ƙarfin PV minti 5 gaba. Ta amfani da samfurin LSTM, shigarwar za ta kasance jerin ma'aunin wutar 10 na ƙarshe da hotunan sama 10. Samfurin yana sarrafa hotuna ta hanyar CNN don fitar da fasalulluka, sannan LSTM ta kama tsarin lokaci. Fitarwa ita ce ƙimar wutar da aka daidaita guda ɗaya, wanda za a iya mayar da ita zuwa ainihin kW. Wannan hasashen yana bawa manajan damar daidaita ayyukan grid, kamar kunna janareto na ajiya ko sarrafa ajiyar makamashi, don kiyaye kwanciyar hankali.
6. Aikace-aikace na Gaba da Hasashe
Tsarin koyon zurfi da aka gabatar yana da hanyoyi masu ban sha'awa da yawa na gaba:
- Haɗin kai da IoT: Tura samfurin akan na'urorin gefe don hasashen lokaci-lokaci, ƙarancin jinkiri a wuraren shigar da hasken rana guda ɗaya.
- Hasashen wurare da yawa: Fadada samfurin don hasashen ƙarfin wutar a fadin gonakin hasken rana da yawa da ke rarrabe ta hanyar amfani da tsarin yanayi na gama gari.
- Samfuran haɗe-haɗe: Haɗa LSTM da hanyoyin hankali ko tsarin mai canzawa don kama dogaro mai nisa a cikin jerin hotunan sama.
- Koyon canja wuri: Horar da CNN a gaba akan manyan tarin hotunan sama (misali, daga tashoshin yanayi) don inganta gama gari zuwa sabbin wurare.
- Ƙididdigar rashin tabbas: Haɗa hanyoyin sadarwa na Bayesian ko Monte Carlo dropout don samar da tazara na hasashen, yana taimakawa yanke shawara mai sanin haɗari.
7. Sharhin Masana
Mahimman Fahimta: Wannan takarda ta nuna cewa koyon zurfi, musamman hanyoyin sadarwa na LSTM, na iya inganta hasashen wutar PV na ɗan gajeren lokaci sosai ta hanyar yin samfurin dogaro na lokaci a cikin hotunan sama da bayanan wutar.
Tsarin Ma'ana: Marubutan sun bayyana matsalar a sarari, sun gabatar da tsarin sadarwa guda uku masu rikitarwa, kuma sun kimanta su bisa tsari akan tarin bayanai na ainihi. Nazarin cirewa ya tabbatar da gudummawar kowane bangare.
Ƙarfi da Rashi: Babban ƙarfin shine kwatankwacin tsarin sadarwa da kuma nuna fifikon LSTM a sarari. Koyaya, tarin bayanai ya iyakance ga wuri ɗaya (Kyoto), kuma ba a bincika aikin samfurin a ƙarƙashin yanayin yanayi mai tsanani (misali, ruwan sama mai yawa, dusar ƙanƙara) ba. Bugu da ƙari, ba a tattauna farashin lissafi na samfurin LSTM ba.
Fahimtar Aiki: Ga masu aiki, samfurin LSTM yana ba da ingantacciyar 21% akan ciwawa, wanda yake da mahimmanci ga sarrafa grid. Don tura wannan, ya kamata a saka hannun jari a cikin tattara bayanai mai yawa (kyamarorin sama da ma'aunin wutar) kuma a yi la'akari da lissafin gefe don hasashen lokaci-lokaci. Aikin gaba ya kamata ya mayar da hankali kan tabbatar da wurare da yawa da juriya ga nau'ikan yanayi daban-daban.
8. Bincike na Asali
Wannan takarda ta ba da gudummawa mai mahimmanci ga fagen hasashen makamashin hasken rana ta hanyar nuna tasirin koyon zurfi don hasashen nan take na matakin minti. Babban sabon abu yana cikin amfani da hanyar sadarwa ta LSTM don kama dogaro na lokaci a cikin hotunan sama da bayanan wutar, samun ƙimar fasaha ta RMSE 21% akan tushen ciwawa. Wannan sakamakon ya dace da manyan abubuwan da ke faruwa a hasashen jerin lokaci, inda tsarin maimaitawa ya nuna fifikon aiki a ayyuka kamar hasashen yanayi da hasashen nauyin makamashi (Hochreiter & Schmidhuber, 1997).
Daga mahangar fasaha, kwatancen tsarin MLP, CNN, da LSTM yana ba da fahimta mai mahimmanci. MLP, wanda ya dogara kawai akan ƙimar wutar tarihi, yana aiki a matsayin tushe mai ƙarfi amma ya kasa kama alamun gani na motsin gajimare. CNN ta inganta akan wannan ta hanyar haɗa hotunan sama, amma tana ɗaukar kowane mataki na lokaci da kansa. Ikon LSTM na yin samfurin jerin yana da mahimmanci, saboda motsin gajimare yana da lokaci a asali. Wannan binciken ya yi daidai da bincike a cikin fahimtar bidiyo, inda ake amfani da 3D CNNs da LSTMs don yin samfurin motsi (Tran et al., 2015).
Koyaya, binciken yana da iyakoki. An tattara bayanan a wuri ɗaya a Kyoto, Japan, wanda ke da takamaiman yanayin yanayi (matsakaici tare da yanayi daban-daban). Ba a gwada iyawar samfurin zuwa wasu yanayi (misali, na wurare masu zafi, bushewa) ba. Bugu da ƙari, takardar ba ta magance farashin lissafi na samfurin LSTM ba, wanda zai iya zama mai hana tura shi akan na'urori masu ƙarancin ƙarfi. Ci gaba na baya-bayan nan a cikin tsarin gine-gine masu nauyi, kamar MobileNets (Howard et al., 2017), za a iya bincika don rage lokacin hasashen.
Wani muhimmin al'amari shine ingancin hotunan sama. Takardar ta yi amfani da ROI kafaffe a kusa da rana, wanda bazai kama gajimare da ke motsawa daga gefe ba. Aikin gaba zai iya haɗa hanyoyin hankali don mayar da hankali kan wuraren da suka dace, kamar yadda aka yi a cikin bayanin hoto (Xu et al., 2015). Bugu da ƙari, ana iya inganta aikin samfurin a ƙarƙashin yanayin canzawa cikin sauri (misali, gajimare masu taruwa) ta amfani da ƙudurin lokaci mafi girma (misali, tazarar daƙiƙa 10).
A ƙarshe, wannan takarda ta ba da tushe mai ƙarfi don hasashen PV na nan take ta hanyar koyon zurfi. Samfurin LSTM shine mafi kyau a sarari, amma tura aiki na zahiri yana buƙatar magance ingancin lissafi da iyawar gama gari. Haɗin wannan hanya tare da tsarin grid mai wayo zai iya haifar da ingantaccen sarrafa makamashi mai sabuntawa da inganci.
9. Manazarta
- Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735-1780.
- Howard, A. G., et al. (2017). MobileNets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861.
- Tran, D., et al. (2015). Learning spatiotemporal features with 3D convolutional networks. Proceedings of the IEEE International Conference on Computer Vision.
- Xu, K., et al. (2015). Show, attend and tell: Neural image caption generation with visual attention. Proceedings of the International Conference on Machine Learning.
- Zhang, J., et al. (2018). Deep photovoltaic nowcasting. Solar Energy (accepted).